Heterogeneous industrial equipment-oriented adaptive data fusion and decision optimization system

Abstract

A heterogeneous industrial equipment-oriented adaptive data fusion and decision optimization system relates to the technical field of industrial automation and information, and comprises a cloud, and the cloud is in communication connection with a data acquisition module, a protocol labeling module, a protocol self-consistent module, a detection decision module, a fast response compensation module and a critical trend detection module; the data acquisition module is used for acquiring multi-source type data of each heterogeneous industrial device, and the protocol labeling module is used for constructing an industrial connection topological graph and performing initial communication protocol labeling on the industrial connection topological graph; the protocol self-consistent module is used for carrying out interactive feature analysis on each node in the industrial connection topological graph and carrying out automatic protocol matching on protocol abnormal nodes; the detection decision module detects the multi-source type data of each node in real time, and selects the rapid response compensation module to execute the rapid response compensation operation or the critical trend detection module to execute the critical trend detection operation according to a monitoring result, so that the production efficiency is remarkably improved, and the product quality is ensured.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation and information technology, and in particular to an adaptive data fusion and decision optimization system for heterogeneous industrial equipment. Background Art

[0002] A Chinese patent with publication number CN118967109B discloses an industrial equipment maintenance method and platform based on multi-source heterogeneous data, including the following steps: obtaining multi-source heterogeneous data, including equipment operation and maintenance data, monitoring data, factory operation data and monitoring image data; generating equipment degradation health indicators based on the operation and maintenance data to quantitatively indicate the health status of the equipment; extracting features from the factory operation data to generate feature changes; constructing a convolutional neural network model, inputting monitoring images into the model, and outputting violation behavior images; performing maintenance operations on factory equipment according to the equipment degradation health indicators, factory operation data feature changes and violation behavior images; A Chinese patent with publication number CN112529184B discloses an industrial process optimization decision-making method that integrates domain knowledge and multi-source data, including: using probabilistic soft logic to obtain industrial process domain knowledge and establish an industrial process domain rule knowledge base; integrating multi-source data semantics and multi-source data features to form a new industrial process semantic knowledge representation and construct an industrial process semantic knowledge base; under the posterior regularization framework, using the industrial process domain rule knowledge base and the semantic knowledge base to obtain an optimization decision model and a posterior distribution model embedded with domain rule knowledge; knowledge distillation technology migrates the knowledge in the optimization decision model embedded with domain rule knowledge to the posterior distribution model.

[0003] Under the current trend of Industry 4.0, industrial production systems are becoming more and more complex, and heterogeneous industrial equipment is widely used in various production links. Equipment produced by different manufacturers has different data interfaces, communication protocols and data formats, which makes it difficult to effectively integrate multi-source data and forms data islands. For example, traditional programmable logic controllers (PLCs) use specific communication protocols for data transmission, while new smart sensors may use wireless communication technology and different data formats, making it extremely challenging to obtain unified and effective data in an industrial production workshop. At the same time, the industrial production environment is complex and changeable. Factors such as the rapid changes in market demand, fluctuations in raw material quality, and aging of equipment all require companies to make accurate and timely decisions based on real-time data to optimize the production process, improve production efficiency, reduce costs, and ensure product quality. However, existing data processing and decision-making systems are often unable to adapt to such dynamic changes. For example, the response delay of optimization strategies based on historical data (such as genetic algorithms) is greater than 30s, and they cannot cope with millisecond-level equipment disturbances (such as sudden changes in injection molding machine pressure). It is difficult to achieve efficient integration of multi-source heterogeneous data and accurate decision optimization. Summary of the invention

[0004] In order to solve the above technical problems, the object of the present invention is to provide an adaptive data fusion and decision optimization system for heterogeneous industrial equipment, including a cloud, wherein the cloud is communicatively connected with a data acquisition module, a protocol annotation module, a protocol self-consistent module, a detection decision module, a rapid response compensation module and a critical trend detection module; The data acquisition module is used to collect multi-source data of various heterogeneous industrial equipment, mark the collection time, and set the collection cycle; The protocol annotation module is used to construct an industrial connection topology map and perform initial communication protocol annotation on the industrial connection topology map; The protocol self-consistent module is used to analyze the interactive characteristics of each node in the industrial connection topology diagram and automatically match the protocol for the protocol abnormal nodes; The detection decision module is used to perform real-time detection of multi-source data of each node, and select to perform rapid response compensation operation or critical trend detection operation according to the monitoring results; The rapid response compensation module is used to generate a temporary compensation model and a joint compensation model, perform a rapid response compensation operation on the abnormal node according to the temporary compensation model, update the joint compensation model according to the rapid response compensation operation result, and then perform a joint compensation operation according to the joint compensation model; The critical trend detection module is used to extract several critical abnormal trend features of each node, perform critical trend detection operations on each normal node, and issue parameters to the rapid response compensation module according to the critical trend detection operation results.

[0005] Furthermore, multi-source data include: equipment operating status data (including vibration signals, temperature data, pressure data, etc.), equipment performance data (including production efficiency data, product quality data, etc.), equipment communication data (including baud rate, data frame interval, CRC check bit), and environmental data (including humidity data, dust concentration data, etc.).

[0006] Furthermore, the protocol annotation module constructs an industrial connection topology map, and the process of initial communication protocol annotation on the industrial connection topology map includes: Obtain assembly connection relationships of several heterogeneous industrial equipment, wherein the several heterogeneous industrial equipment include terminal equipment such as sensors, actuators, motors, valves, edge control centers, human-computer interaction terminals, edge computing nodes, etc. in the production line. The assembly connection relationships of the several heterogeneous industrial equipment include direct connection between heterogeneous terminal equipment, connection relationship between terminal equipment and edge control centers (such as PLC integrated MQTT client and other protocol conversion gateways), connection relationship between terminal equipment and human-computer interaction terminals (touch screen, industrial computer, large screen in central control room), connection relationship between terminal equipment and edge computing nodes (edge ​​server, gateway), etc. Take several heterogeneous industrial equipment as nodes of an industrial connection topology map, and the nodes are used to store multi-source type data of each heterogeneous industrial equipment. Take the assembly connection relationship of several heterogeneous industrial equipment as the connection relationship between nodes, construct an industrial connection topology map, obtain the initial communication protocol between several heterogeneous industrial equipment, and mark the connection relationship between each node with the initial communication protocol.

[0007] Furthermore, the protocol self-consistent module performs interactive feature analysis on each node in the industrial connection topology diagram, and the process of automatically matching the protocol with abnormal nodes includes: According to the initial communication protocol between the nodes, the threshold intervals of various types of interaction indicators between the nodes are preset, and the interaction features of the nodes with connection relationships are extracted. According to the multi-source type data between the nodes with connection relationships, various types of interaction indicators between the nodes with connection relationships are obtained. The types of interaction indicators include frequent data loss, long data delay, parsing success rate, connection establishment time, disconnection and reconnection times, function execution results, etc. Each type of interaction indicator is compared with the corresponding interaction indicator threshold interval. If the type of interaction indicator is not within the corresponding interaction indicator threshold interval, the node with connection relationship is marked as a protocol abnormal node; Extract protocol fingerprint features from protocol abnormal nodes, obtain protocol fingerprint features, build a protocol fingerprint atlas library, and use reinforcement learning to build an automatic protocol matching model based on the protocol fingerprint atlas library; The process of building a protocol fingerprint library includes: building a control rule table of an automatic protocol matching model according to the protocol fingerprint library, defining the state of the automatic protocol matching model as the protocol fingerprint features of the currently collected node. Including the features of baud rate, data frame interval and cyclic redundancy check (CRC check bit), for example, the state s=[9600,10ms,CRC16] means that the baud rate of the currently collected data is 9600, the data frame interval is 10ms, and the CRC16 check algorithm is used. The action is defined as a possible protocol type, such as Modbus, OPCUA, Profibus, etc., and a reward function is set. The reward function is used to evaluate the quality of an action taken by the agent in a certain state. If the agent correctly matches the protocol type, a positive reward is given. If the match is wrong, a negative reward is given. The reinforcement learning algorithm continuously interacts with the environment, updates the control rule table, obtains the experience score of each state-action combination, and thus finds the optimal action strategy. When new data of the node is collected, the protocol fingerprint feature of the data is input into the control rule value table as the state, and the action with the largest experience score is selected as the matching protocol type; The protocol fingerprint feature is input into the automatic protocol matching model, the communication protocol of the protocol abnormal node is output according to the automatic protocol matching model, and the initial communication protocol of the protocol abnormal node is updated according to the communication protocol.

[0008] Furthermore, the process of constructing the fingerprint protocol graph includes: Based on the idea of ​​simulation learning, a simulator is used to generate interactive training data of different communication protocols. Baud rate analysis, data frame interval analysis and cyclic redundancy check analysis are performed on the interactive training data of different communication protocols. The baud rate probability model, data frame interval probability model and cyclic redundancy check probability model are constructed. The baud rate probability model, data frame interval probability model and cyclic redundancy check probability model are combined to construct a fingerprint protocol map.

[0009] Furthermore, baud rate refers to the speed of data transmission, and different industrial protocols may use different baud rates. The probability distribution of baud rate can be constructed by statistically analyzing the baud rate usage of a large number of known protocols. For example, if 1,000 devices using the Modbus protocol are collected and it is found that 800 of them use 9600 baud rate, then the probability of 9600 baud rate in the Modbus protocol is 0.8.

[0010] The data frame interval refers to the time interval between two adjacent data frames. Different protocols have different time rules for sending and receiving data frames, so the data frame interval can also be used as a feature of the protocol. The data frame intervals of different protocols can be sampled and counted to obtain the probability distribution of the data frame interval of each protocol.

[0011] Cyclic redundancy check (CRC) is used to detect errors in data transmission. Different protocols may use different CRC algorithms and parameters, so the characteristics of CRC check bits can also be used to distinguish protocols. By analyzing the CRC generator polynomials, initial values ​​and other parameters of different protocols, a probability model of CRC check bits is constructed.

[0012] By combining the probability models of baud rate, data frame interval and CRC check bit, we can construct a protocol fingerprint map. Each protocol has a corresponding fingerprint vector, and each element in the vector represents the probability of the protocol on a certain feature. For example, for a protocol fingerprint vector [0.8, 0.6, 0.7], it means that the probability of the protocol using 9600 baud rate is 0.8, the probability of the data frame interval being within a certain range is 0.6, and the probability of using a certain CRC algorithm is 0.7.

[0013] Furthermore, the detection decision module performs real-time detection on multi-source data of each node, and the process of selecting to perform a rapid response compensation operation or a critical trend detection operation according to the monitoring results includes: Extract the numerical time series sequence of each type of indicator in the multi-source type data of each node in the current collection cycle, preset the standard threshold interval corresponding to each type of indicator of each node, compare the numerical time series sequence of each type of indicator with the corresponding standard threshold interval, and obtain the cumulative time that each type of indicator is not in the corresponding standard threshold interval; A cumulative time threshold is preset. If the cumulative time of a type indicator is greater than the cumulative time threshold, a sudden abnormal alarm signal is generated and a rapid response compensation operation is performed. If the cumulative time of each type of indicator is less than or equal to the cumulative time threshold, the node is marked as a normal node and a critical trend detection operation is performed.

[0014] Furthermore, the process of generating a temporary compensation model by the rapid response compensation module includes: Mark the corresponding type indicator whose cumulative time is greater than the cumulative time threshold as a key indicator, mark the node to which the key indicator belongs as an abnormal node, and extract features from the multi-source type data of the abnormal node in the current collection cycle based on the key indicator to obtain a feature vector set. Taking the sudden change in bearing temperature as an example, extract the feature vector set related to the sudden change in bearing temperature from the multi-source type data, including the temperature change rate (such as the temperature change every 10 seconds), the main frequency of the vibration signal, the harmonic distortion rate of the current, etc. These features can more accurately reflect the abnormal state of the equipment; Construct a sudden abnormal fault library, which includes several historical fault cases, which are classified and clustered according to equipment type, fault type, related features, etc. For example, for bearing fault cases, they may be clustered according to "temperature mutation + vibration main frequency" to form different fault categories, and compare the feature data set with the cases in the sudden abnormal fault library by cosine similarity to obtain the cosine similarity between the current feature vector set and the feature vector set of historical fault cases, and select the top k historical fault cases with the highest cosine similarity; A temporary compensation model is constructed based on a decision tree model. The temporary compensation model constructed based on the decision tree model has a short inference time (usually less than 10 milliseconds), which is suitable for fast operation on the edge side. The compensation strategies of the top k historical fault cases are extracted to customize the parameters of the temporary compensation model, and the temporary compensation model is verified based on the multi-source data of abnormal nodes in the current acquisition cycle. For example, if the compensation strategy for excessive bearing temperature in similar cases adopts a compensation strategy of reducing speed by 15% and increasing lubrication flow by 20%, and increasing the speed reduction by 2% for every 1°C increase, then the corresponding compensation parameters are also injected into the temporary model. Then, in the model verification process, first The compensation strategy of the case with the highest cosine similarity is selected from the first k historical fault cases to customize the parameters of the temporary compensation model. The customized model is then verified using the multi-source data of the abnormal node in the current acquisition cycle. The operating status of the equipment (such as temperature change) is predicted by the model and compared with the actual situation. If the deviation between the predicted result and the actual situation exceeds the preset deviation threshold, the compensation strategy and related parameters of the case with the second highest cosine similarity are selected from the first k historical fault cases to customize the parameters of the temporary compensation model. The above process is repeated until the deviation between the predicted result and the actual situation does not exceed the preset deviation threshold, and the verified temporary compensation model is obtained. Furthermore, the rapid response compensation module generates a joint compensation model, performs a rapid response compensation operation on the abnormal node according to the temporary compensation model, and updates the joint compensation model according to the result of the rapid response compensation operation, including: Construct an association rule knowledge graph according to the industrial connection topology graph, wherein the association rule knowledge graph includes multi-source type data, spatiotemporal feature sequences, and joint compensation strategies between interconnected nodes, construct a joint compensation model based on deep learning, use the multi-source type data, spatiotemporal feature sequences, and joint compensation strategies as training sets and test sets, input the training sets into the joint compensation model for training until the loss function training is stable, save the model parameters, test the joint compensation model through the test set, and output the joint compensation model until it meets the preset requirements; Building a joint compensation model based on deep learning is a complex process, which involves multiple steps such as model selection, training, verification and testing. The following is a detailed supplementary explanation of this process: The present invention selects RBF neural network as the deep learning architecture. After determining the model architecture, cross entropy loss (Binary Cross-Entropy Loss) is selected as the optimization target, and then the prepared training set is input into the selected deep learning model to start training. During the training process, the weights are continuously updated through the back propagation algorithm, so that the loss function gradually decreases until a stable state is reached. During this period, the early stopping method and other techniques are used to avoid overfitting. In addition to the basic training process, various parameters of the model are tuned by grid search, including learning rate, batch size, regularization coefficient, etc.

[0015] After the model training is completed and the parameters are adjusted, the final evaluation is performed through the test set to obtain the evaluation results of the model, which include classification indicators such as accuracy, recall, F1 score, etc. Based on the evaluation results on the test set, it is judged whether the model has reached the expected standards. If the requirements are met, the model parameters are saved and ready for deployment; if not ideal, it is necessary to return to a previous stage to re-examine issues such as data quality, model structure or training strategy.

[0016] Furthermore, the spatiotemporal feature sequences between interconnected nodes, for example: a sudden change in the bearing temperature of device A (ΔT>5°C / 10s) will cause a drop in the outlet pressure of the lubrication pump of device B (ΔP<-0.2MPa). Based on the device dependency relationship in the industrial connection topology (such as the oil connection between the bearing and the lubrication pump), it is determined that insufficient lubrication of device B is the main cause of the temperature rise of device A, and the wear of the lubrication pump plunger is locked; Joint compensation strategies include: Edge side: instant stop loss (<10ms), for example, device A: based on the temporary compensation model, immediately perform local load reduction (such as reducing the rolling mill speed from 1200rpm to 900rpm to reduce friction heat); device B: trigger the emergency mode of the lubrication pump based on the temporary compensation model (the plunger frequency is increased from 50Hz to 60Hz, and short-term overload is allowed for 10 minutes); Workshop level: Collaborative adjustment (<5s). For example, the cloud sends a linkage strategy to the workshop controller through the OPC UA protocol: start the backup lubrication pump C (ready within 30s), and reduce the load of device A to 50% (to avoid overloading a single pump); adjust the cooling water flow rate of the adjacent device D (to compensate for the imbalance of the production line rhythm caused by the load reduction of device A); The temporary compensation model is packaged and deployed. The specific process of model packaging and deployment includes: packaging the customized model into an executable file, such as the TensorFlow Lite micromodel. This model file is small in size (such as less than 100KB) and is suitable for running on edge devices. Define an interface for the packaged temporary compensation model. For example, subscribe to the Topic of bearing temperature through the MQTT protocol to obtain temperature data in real time, and write the compensation instructions output by the model to the corresponding registers through the Modbus protocol to achieve control of industrial equipment. Finally, set a validity period for the temporary model, such as automatically marking "valid for 2 minutes". During the validity period, the model continues to run and compensates the equipment; after the validity period, the model automatically becomes invalid to avoid outdated compensation strategies from causing adverse effects on the equipment. The abnormal node is compensated according to the temporary compensation model, and the compensation time threshold is preset. When the abnormal node is marked as a normal node within the compensation time threshold, the compensation parameters and effective data of the temporary compensation model are collected to update the association rule knowledge graph, and the joint compensation model is incrementally trained according to the updated association rule knowledge graph. For example, the new association rule of "bearing temperature rise-speed reduction compensation" is added to provide richer knowledge support for subsequent fault diagnosis and processing. If the abnormal node is not marked as a normal node within the compensation time threshold, the operation of the abnormal node is suspended, and the heterogeneous industrial equipment of the abnormal node is inspected and repaired.

[0017] Furthermore, the process of the rapid response compensation module performing the joint compensation operation according to the joint compensation model includes: When an abnormal node is marked as a normal node, time features and spatial features are extracted from multi-source type data of the normal node and other normal nodes connected to the normal node in the current acquisition cycle to obtain a spatiotemporal feature sequence. The multi-source type data and the spatiotemporal feature sequence are input into the joint compensation model after incremental training, and a joint compensation strategy is output according to the joint compensation model.

[0018] Furthermore, the process of extracting temporal and spatial features from multi-source data and obtaining temporal and spatial feature sequences includes: Extract the numerical time series of each type of indicator corresponding to the multi-source type data, input the numerical time series of each type of indicator into the time convolutional neural network, and obtain the numerical change trend characteristics of each type of indicator according to the trained time convolutional network; The industrial connection topology graph is learned through the graph attention network, and the numerical change trend characteristics of various types of indicators of normal nodes and other nodes that have a connection relationship with normal nodes are imported into the graph attention network. The attention weights of other normal nodes to the target normal node are obtained through the attention mechanism, and the attention weights are distributed to each adjacent normal node. The aggregate representation of the generated features is used to generate the spatiotemporal feature sequence of multi-source type data; Furthermore, the attention weights of the adjacent normal node j and the target normal node i are obtained The calculation formula is: ; in, () represents the activation function, which is used to introduce nonlinear characteristics. Represents the feature vector and The splicing operation, represents a learnable vector, represents a learnable weight matrix, represents the feature vector of the target normal node i, Represents the feature vector of the adjacent normal node j; By the attention value between the target normal node i and the adjacent normal node j The attention weight is determined by the ratio of the sum of the attention values ​​of the target normal node i and all adjacent normal nodes. After normalizing the attention weights, the features of the target normal nodes are updated using the neighbor aggregation mechanism. ; in, represents the final updated representation of the target normal node i, represents the activation function, G represents the parameter matrix of feature transformation, j represents the number of adjacent normal node items, Represents the set of adjacent normal nodes of the target normal node i.

[0019] Furthermore, the process of extracting several critical abnormal trend features of each node by the critical trend detection module includes: Extract the moment when the node is marked as an abnormal node in the historical collection period, extract the numerical time series sequence and key indicators of various types of indicators between the start timestamp of the historical collection period and the moment, mark other key indicators other than the key indicators as non-key indicators, mark the numerical time series sequence as a critical abnormal numerical time series sequence, and mark the historical collection period as a critical abnormal collection period; Obtaining a standard time series sequence of key indicators within a critical anomaly acquisition period, performing an autocorrelation comparison between the critical anomaly value time series sequence of the key indicators and the standard time series sequence, obtaining an autocorrelation coefficient of the key indicators, performing a cross-correlation comparison between the critical anomaly value time series sequences between the key indicators, obtaining a mutual correlation coefficient between the key indicators, performing a cross-correlation comparison between the critical anomaly value time series sequences of the key indicators and non-key indicators, obtaining a mutual correlation coefficient between the key indicators and the non-key indicators, constructing a critical anomaly trend feature according to the autocorrelation coefficient of the key indicators, the mutual correlation coefficient between the key indicators, and the mutual correlation coefficient between the key indicators and the non-key indicators, and extracting a temporary compensation model for the critical anomaly acquisition period, and associating the temporary compensation model with the critical anomaly trend feature; A trend feature database is constructed, and critical abnormal trend features within several abnormal collection cycles of each node are stored in the trend feature database.

[0020] Furthermore, the process of performing autocorrelation comparison includes: ; in, represents the autocorrelation coefficient of type index i, represents the value of type i at time t, represents the value of the standard time series of type index i at time t, and n represents the total number of time points; The process of performing a cross-correlation comparison includes: ; in, represents the mutual correlation coefficient between type index i and type index j; Represents the value of type index i at time t; represents the average value of type index i; represents the value of type index j at time t; Represents the average value of type index j.

[0021] Furthermore, the process of obtaining the standard time series of each type of indicator within the historical collection period includes: A data prediction model is constructed based on deep learning. The multi-source data of each node in several historical collection cycles are used as training data. The data prediction model is trained using the training data to obtain the trained data prediction model. The multi-source data in the critical anomaly collection cycle are input into the data prediction model, and the standard time series of each type of indicator is output according to the data prediction model.

[0022] Furthermore, the critical trend detection module performs a critical trend detection operation on each normal node, and according to the critical trend detection operation result, performs a parameter delivery operation on the rapid response compensation module, including: The trend feature is extracted from the numerical time series of each type of indicator in the current acquisition cycle of the normal node, and the trend feature in the current acquisition cycle of the normal node (including the autocorrelation coefficient of each type of indicator and the mutual correlation coefficient between each type of indicator) is obtained. According to the trend feature, a search and comparison is performed in the trend feature database to obtain the similarity between the trend feature and several critical abnormal trend features. If the similarity between the trend feature and the critical abnormal trend feature is greater than the preset similarity threshold, the normal node is marked as a critical warning node, and the parameters of the temporary compensation model of the critical warning node are customized according to the temporary compensation model associated with the critical abnormal trend feature. The process of screening out the top k historical fault cases with the highest cosine similarity is directly skipped, and then a rapid response compensation operation is performed on the critical warning node according to the temporary compensation model with customized parameters.

[0023] Compared with the prior art, the present invention has the following beneficial effects: 1. The protocol annotation module constructs the industrial connection topology diagram and annotates the initial communication protocol, so that the system can clearly grasp the connection relationship and initial protocol between devices. The protocol self-consistent module analyzes the interaction characteristics of each node and automatically matches the protocol, which can timely discover and solve protocol anomalies and ensure smooth and stable communication between devices. It avoids data transmission interruptions or errors caused by protocol incompatibility and improves the compatibility and reliability of the system.

[0024] 2. The detection decision module detects multi-source data of each node in real time, and selects to execute rapid response compensation operation or critical trend detection operation according to the monitoring results, realizing intelligent judgment and targeted processing of equipment status. It can timely discover sudden abnormalities and potential critical trends of equipment, take corresponding measures to avoid equipment failure and reduce production losses.

[0025] 3. The rapid response compensation module generates temporary compensation models and joint compensation models, performs rapid response compensation operations on abnormal nodes, and updates the joint compensation model based on the operation results. This rapid response mechanism can take quick measures when equipment abnormalities occur, reduce the impact of abnormalities on production, and continuously update the joint compensation model to improve the system's ability to handle similar abnormalities, and enhance the system's adaptability and self-learning ability.

[0026] 4. The critical trend detection module extracts the critical abnormal trend characteristics of each node, performs critical trend detection operations on normal nodes, and issues parameters to the rapid response compensation module based on the detection results. Generate trend characteristics based on the multi-source data of the current acquisition cycle, retrieve and compare the trend characteristics in the trend characteristic database, and generate preventive measures based on the comparison results, realizing the prediction of early faults before the industrial equipment changes from a normal state to a fault state, so as to take preventive measures before the industrial equipment fault matures, improving the reliability of the equipment and the continuity of production. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of an adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following is a clear and complete description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0029] like Figure 1 As shown, an adaptive data fusion and decision optimization system for heterogeneous industrial equipment includes a cloud, and the cloud is connected to a data acquisition module, a protocol annotation module, a protocol self-consistent module, a detection decision module, a rapid response compensation module and a critical trend detection module; The data acquisition module is used to collect multi-source data of various heterogeneous industrial equipment, mark the collection time, and set the collection cycle; The protocol annotation module is used to construct an industrial connection topology map and perform initial communication protocol annotation on the industrial connection topology map; The protocol self-consistent module is used to analyze the interactive characteristics of each node in the industrial connection topology diagram and automatically match the protocol for the protocol abnormal nodes; The detection decision module is used to perform real-time detection of multi-source data of each node, and select to perform rapid response compensation operation or critical trend detection operation according to the monitoring results; The rapid response compensation module is used to generate a temporary compensation model and a joint compensation model, perform a rapid response compensation operation on the abnormal node according to the temporary compensation model, update the joint compensation model according to the rapid response compensation operation result, and then perform a joint compensation operation according to the joint compensation model; The critical trend detection module is used to extract several critical abnormal trend features of each node, perform critical trend detection operations on each normal node, and issue parameters to the rapid response compensation module according to the critical trend detection operation results.

[0030] It should be further explained that, in the specific implementation process, the multi-source data include: equipment operation status data (including vibration signals, temperature data, pressure data, etc.), equipment performance data (including production efficiency data, product quality data, etc.), equipment communication data (including baud rate, data frame interval, CRC check bit), environmental data (including humidity data, dust concentration data, etc.).

[0031] It should be further explained that, in the specific implementation process, the protocol annotation module constructs the industrial connection topology map, and the process of initial communication protocol annotation of the industrial connection topology map includes: Obtain assembly connection relationships of several heterogeneous industrial equipment, wherein the several heterogeneous industrial equipment include terminal equipment such as sensors, actuators, motors, valves, edge control centers, human-computer interaction terminals, edge computing nodes, etc. in the production line. The assembly connection relationships of the several heterogeneous industrial equipment include direct connection between heterogeneous terminal equipment, connection relationship between terminal equipment and edge control centers (such as PLC integrated MQTT client and other protocol conversion gateways), connection relationship between terminal equipment and human-computer interaction terminals (touch screen, industrial computer, large screen in central control room), connection relationship between terminal equipment and edge computing nodes (edge ​​server, gateway), etc. Take several heterogeneous industrial equipment as nodes of an industrial connection topology map, and the nodes are used to store multi-source type data of each heterogeneous industrial equipment. Take the assembly connection relationship of several heterogeneous industrial equipment as the connection relationship between nodes, construct an industrial connection topology map, obtain the initial communication protocol between several heterogeneous industrial equipment, and mark the connection relationship between each node with the initial communication protocol.

[0032] It should be further explained that, in the specific implementation process, the protocol self-consistent module analyzes the interactive characteristics of each node in the industrial connection topology diagram, and the process of automatically matching the protocol with abnormal nodes includes: According to the initial communication protocol between the nodes, the threshold intervals of various types of interaction indicators between the nodes are preset, and the interaction features of the nodes with connection relationships are extracted. According to the multi-source type data between the nodes with connection relationships, various types of interaction indicators between the nodes with connection relationships are obtained. The types of interaction indicators include frequent data loss, long data delay, parsing success rate, connection establishment time, disconnection and reconnection times, function execution results, etc. Each type of interaction indicator is compared with the corresponding interaction indicator threshold interval. If the type of interaction indicator is not within the corresponding interaction indicator threshold interval, the node with connection relationship is marked as a protocol abnormal node; Extract protocol fingerprint features from protocol abnormal nodes, obtain protocol fingerprint features, build a protocol fingerprint atlas library, and use reinforcement learning to build an automatic protocol matching model based on the protocol fingerprint atlas library; The process of building a protocol fingerprint library includes: building a control rule table of an automatic protocol matching model according to the protocol fingerprint library, defining the state of the automatic protocol matching model as the protocol fingerprint features of the currently collected node. Including the features of baud rate, data frame interval and cyclic redundancy check (CRC check bit), for example, the state s=[9600,10ms,CRC16] means that the baud rate of the currently collected data is 9600, the data frame interval is 10ms, and the CRC16 check algorithm is used. The action is defined as a possible protocol type, such as Modbus, OPCUA, Profibus, etc., and a reward function is set. The reward function is used to evaluate the quality of an action taken by the agent in a certain state. If the agent correctly matches the protocol type, a positive reward is given. If the match is wrong, a negative reward is given. The reinforcement learning algorithm continuously interacts with the environment, updates the control rule table, obtains the experience score of each state-action combination, and thus finds the optimal action strategy. When new data of the node is collected, the protocol fingerprint feature of the data is input into the control rule value table as the state, and the action with the largest experience score is selected as the matching protocol type; The protocol fingerprint feature is input into the automatic protocol matching model, the communication protocol of the protocol abnormal node is output according to the automatic protocol matching model, and the initial communication protocol of the protocol abnormal node is updated according to the communication protocol.

[0033] It should be further explained that, in the specific implementation process, the process of constructing the fingerprint protocol map includes: Based on the idea of ​​simulation learning, a simulator is used to generate interactive training data of different communication protocols. Baud rate analysis, data frame interval analysis and cyclic redundancy check analysis are performed on the interactive training data of different communication protocols. The baud rate probability model, data frame interval probability model and cyclic redundancy check probability model are constructed. The baud rate probability model, data frame interval probability model and cyclic redundancy check probability model are combined to construct a fingerprint protocol map.

[0034] It should be further explained that baud rate refers to the speed of data transmission, and different industrial protocols may use different baud rates. The probability distribution of baud rate can be constructed by statistically analyzing the baud rate usage of a large number of known protocols. For example, if 1000 devices using the Modbus protocol are collected and 800 of them use 9600 baud rate, then the probability of 9600 baud rate in the Modbus protocol is 0.8.

[0035] The data frame interval refers to the time interval between two adjacent data frames. Different protocols have different time rules for sending and receiving data frames, so the data frame interval can also be used as a feature of the protocol. The data frame intervals of different protocols can be sampled and counted to obtain the probability distribution of the data frame interval of each protocol.

[0036] Cyclic redundancy check (CRC) is used to detect errors in data transmission. Different protocols may use different CRC algorithms and parameters, so the characteristics of CRC check bits can also be used to distinguish protocols. By analyzing the CRC generator polynomials, initial values ​​and other parameters of different protocols, a probability model of CRC check bits is constructed.

[0037] By combining the probability models of baud rate, data frame interval and CRC check bit, we can construct a protocol fingerprint map. Each protocol has a corresponding fingerprint vector, and each element in the vector represents the probability of the protocol on a certain feature. For example, for a protocol fingerprint vector [0.8, 0.6, 0.7], it means that the probability of the protocol using 9600 baud rate is 0.8, the probability of the data frame interval being within a certain range is 0.6, and the probability of using a certain CRC algorithm is 0.7.

[0038] It should be further explained that, in the specific implementation process, the detection decision module performs real-time detection on multi-source data of each node, and the process of selecting to perform a rapid response compensation operation or a critical trend detection operation according to the monitoring results includes: Extract the numerical time series sequence of each type of indicator in the multi-source type data of each node in the current collection cycle, preset the standard threshold interval corresponding to each type of indicator of each node, compare the numerical time series sequence of each type of indicator with the corresponding standard threshold interval, and obtain the cumulative time that each type of indicator is not in the corresponding standard threshold interval; A cumulative time threshold is preset. If the cumulative time of a type indicator is greater than the cumulative time threshold, a sudden abnormal alarm signal is generated and a rapid response compensation operation is performed. If the cumulative time of each type of indicator is less than or equal to the cumulative time threshold, the node is marked as a normal node and a critical trend detection operation is performed.

[0039] It should be further explained that, in the specific implementation process, the process of the rapid response compensation module generating a temporary compensation model includes: Mark the corresponding type indicator whose cumulative time is greater than the cumulative time threshold as a key indicator, mark the node to which the key indicator belongs as an abnormal node, and extract features from the multi-source type data of the abnormal node in the current collection cycle based on the key indicator to obtain a feature vector set. Taking the sudden change in bearing temperature as an example, extract the feature vector set related to the sudden change in bearing temperature from the multi-source type data, including the temperature change rate (such as the temperature change every 10 seconds), the main frequency of the vibration signal, the harmonic distortion rate of the current, etc. These features can more accurately reflect the abnormal state of the equipment; Construct a sudden abnormal fault library, which includes several historical fault cases, which are classified and clustered according to equipment type, fault type, related features, etc. For example, for bearing fault cases, they may be clustered according to "temperature mutation + vibration main frequency" to form different fault categories, and compare the feature data set with the cases in the sudden abnormal fault library by cosine similarity to obtain the cosine similarity between the current feature vector set and the feature vector set of historical fault cases, and select the top k historical fault cases with the highest cosine similarity; A temporary compensation model is constructed based on a decision tree model. The temporary compensation model constructed based on the decision tree model has a short inference time (usually less than 10 milliseconds), which is suitable for fast operation on the edge side. The compensation strategies of the top k historical fault cases are extracted to customize the parameters of the temporary compensation model, and the temporary compensation model is verified based on the multi-source data of abnormal nodes in the current acquisition cycle. For example, if the compensation strategy for excessive bearing temperature in similar cases adopts a compensation strategy of reducing speed by 15% and increasing lubrication flow by 20%, and increasing the speed reduction by 2% for every 1°C increase, then the corresponding compensation parameters are also injected into the temporary model. Then, in the model verification process, first The compensation strategy of the case with the highest cosine similarity is selected from the first k historical fault cases to customize the parameters of the temporary compensation model. The customized model is then verified using the multi-source data of the abnormal node in the current acquisition cycle. The operating status of the equipment (such as temperature change) is predicted by the model and compared with the actual situation. If the deviation between the predicted result and the actual situation exceeds the preset deviation threshold, the compensation strategy and related parameters of the case with the second highest cosine similarity are selected from the first k historical fault cases to customize the parameters of the temporary compensation model. The above process is repeated until the deviation between the predicted result and the actual situation does not exceed the preset deviation threshold, and the verified temporary compensation model is obtained. The temporary compensation model on the edge side is generated and executed locally. The output of the compensation instruction can be controlled within 10ms from abnormal detection (such as through pre-stored lightweight model templates + real-time data matching), directly driving the device to execute (such as immediate speed reduction, increased lubrication). In the factory environment, network interruptions caused by wireless signal obstruction, switch failure, etc. occur frequently. If the joint compensation model is relied on, the device will lose its protection capability when the network is disconnected; The temporary compensation model runs offline and relies on historical cases and model templates stored on the edge to ensure that compensation operations can be completed even if the network is disconnected; At the same time, parameter fine-tuning based on similar cases is more in line with the current equipment characteristics than general models (such as differences in bearing compensation under different lubrication conditions), and it has its own "self-destruction mechanism" to avoid becoming a new source of failure. At the same time, through knowledge feedback, the system becomes smarter and smarter. This design breaks through the traditional passive mode of "failure means shutdown" and realizes the key transformation of industrial AI from "post-analysis" to "real-time life-saving".

[0040] It should be further explained that, in the specific implementation process, the rapid response compensation module generates a joint compensation model, performs a rapid response compensation operation on the abnormal node according to the temporary compensation model, and updates the joint compensation model according to the result of the rapid response compensation operation, including: Construct an association rule knowledge graph according to the industrial connection topology graph, wherein the association rule knowledge graph includes multi-source type data, spatiotemporal feature sequences, and joint compensation strategies between interconnected nodes, construct a joint compensation model based on deep learning, use the multi-source type data, spatiotemporal feature sequences, and joint compensation strategies as training sets and test sets, input the training sets into the joint compensation model for training until the loss function training is stable, save the model parameters, test the joint compensation model through the test set, and output the joint compensation model until it meets the preset requirements; Building a joint compensation model based on deep learning is a complex process, which involves multiple steps such as model selection, training, verification and testing. The following is a detailed supplementary explanation of this process: The present invention selects RBF neural network as the deep learning architecture. After determining the model architecture, cross entropy loss (Binary Cross-Entropy Loss) is selected as the optimization target, and then the prepared training set is input into the selected deep learning model to start training. During the training process, the weights are continuously updated through the back propagation algorithm, so that the loss function gradually decreases until a stable state is reached. During this period, the early stopping method and other techniques are used to avoid overfitting. In addition to the basic training process, various parameters of the model are tuned by grid search, including learning rate, batch size, regularization coefficient, etc.

[0041] After the model training is completed and the parameters are adjusted, the final evaluation is performed through the test set to obtain the evaluation results of the model, which include classification indicators such as accuracy, recall, F1 score, etc. Based on the evaluation results on the test set, it is judged whether the model has reached the expected standards. If the requirements are met, the model parameters are saved and ready for deployment; if not ideal, it is necessary to return to a previous stage to re-examine issues such as data quality, model structure or training strategy.

[0042] It should be further explained that the time-space characteristic sequence between interconnected nodes, for example: a sudden change in the bearing temperature of device A (ΔT>5℃ / 10s) will cause a drop in the outlet pressure of the lubrication pump of device B (ΔP<-0.2MPa). Based on the device dependency relationship in the industrial connection topology (such as the oil connection between the bearing and the lubrication pump), it is determined that insufficient lubrication of device B is the main cause of the temperature rise of device A, and the wear of the lubrication pump plunger is locked; Joint compensation strategies include: Edge side: instant stop loss (<10ms), for example, device A: based on the temporary compensation model, immediately perform local load reduction (such as reducing the rolling mill speed from 1200rpm to 900rpm to reduce friction heat); device B: trigger the emergency mode of the lubrication pump based on the temporary compensation model (the plunger frequency is increased from 50Hz to 60Hz, and short-term overload is allowed for 10 minutes); Workshop level: Collaborative adjustment (<5s). For example, the cloud sends a linkage strategy to the workshop controller through the OPC UA protocol: start the backup lubrication pump C (ready within 30s), and reduce the load of device A to 50% (to avoid overloading a single pump); adjust the cooling water flow rate of the adjacent device D (to compensate for the imbalance of the production line rhythm caused by the load reduction of device A); The temporary compensation model is packaged and deployed. The specific process of model packaging and deployment includes: packaging the customized model into an executable file, such as the TensorFlow Lite micromodel. This model file is small in size (such as less than 100KB) and is suitable for running on edge devices. Define an interface for the packaged temporary compensation model. For example, subscribe to the Topic of bearing temperature through the MQTT protocol to obtain temperature data in real time, and write the compensation instructions output by the model to the corresponding registers through the Modbus protocol to achieve control of industrial equipment. Finally, set a validity period for the temporary model, such as automatically marking "valid for 2 minutes". During the validity period, the model continues to run and compensates the equipment; after the validity period, the model automatically becomes invalid to avoid outdated compensation strategies from causing adverse effects on the equipment. The abnormal node is compensated according to the temporary compensation model, and the compensation time threshold is preset. When the abnormal node is marked as a normal node within the compensation time threshold, the compensation parameters and effective data of the temporary compensation model are collected to update the association rule knowledge graph, and the joint compensation model is incrementally trained according to the updated association rule knowledge graph. For example, the new association rule of "bearing temperature rise-speed reduction compensation" is added to provide richer knowledge support for subsequent fault diagnosis and processing. If the abnormal node is not marked as a normal node within the compensation time threshold, the operation of the abnormal node is suspended, and the heterogeneous industrial equipment of the abnormal node is inspected and repaired.

[0043] It should be further explained that, in the specific implementation process, the process of the rapid response compensation module performing the joint compensation operation according to the joint compensation model includes: When an abnormal node is marked as a normal node, time features and spatial features are extracted from multi-source type data of the normal node and other normal nodes connected to the normal node in the current acquisition cycle to obtain a spatiotemporal feature sequence. The multi-source type data and the spatiotemporal feature sequence are input into the joint compensation model after incremental training, and a joint compensation strategy is output according to the joint compensation model.

[0044] It should be further explained that, in the specific implementation process, the process of extracting temporal and spatial features from multi-source data and obtaining spatiotemporal feature sequences includes: Extract the numerical time series of each type of indicator corresponding to the multi-source type data, input the numerical time series of each type of indicator into the time convolutional neural network, and obtain the numerical change trend characteristics of each type of indicator according to the trained time convolutional network; The industrial connection topology graph is learned through the graph attention network, and the numerical change trend characteristics of various types of indicators of normal nodes and other nodes connected to normal nodes are imported into the graph attention network. The attention weights of other normal nodes to the target normal node are obtained through the attention mechanism, and the attention weights are distributed to each adjacent normal node. The aggregate representation of the generated features is used to generate the spatiotemporal feature sequence of multi-source data. It should be further explained that in the specific implementation process, the attention weights of the adjacent normal node j and the target normal node i are obtained. The calculation formula is: ; in, () represents the activation function, which is used to introduce nonlinear characteristics. Represents the feature vector and The splicing operation, represents a learnable vector, represents a learnable weight matrix, represents the feature vector of the target normal node i, Represents the feature vector of the adjacent normal node j; By the attention value between the target normal node i and the adjacent normal node j The attention weight is determined by the ratio of the sum of the attention values ​​of the target normal node i and all adjacent normal nodes. After normalizing the attention weights, the features of the target normal nodes are updated using the neighbor aggregation mechanism. ; in, represents the final updated representation of the target normal node i, represents the activation function, G represents the parameter matrix of feature transformation, j represents the number of adjacent normal node items, Represents the set of adjacent normal nodes of the target normal node i.

[0045] It should be further explained that, in the specific implementation process, the process of extracting several critical abnormal trend features of each node by the critical trend detection module includes: Extract the moment when the node is marked as an abnormal node in the historical collection period, extract the numerical time series sequence and key indicators of various types of indicators between the start timestamp of the historical collection period and the moment, mark other key indicators other than the key indicators as non-key indicators, mark the numerical time series sequence as a critical abnormal numerical time series sequence, and mark the historical collection period as a critical abnormal collection period; Obtaining a standard time series sequence of key indicators within a critical anomaly acquisition period, performing an autocorrelation comparison between the critical anomaly value time series sequence of the key indicators and the standard time series sequence, obtaining an autocorrelation coefficient of the key indicators, performing a cross-correlation comparison between the critical anomaly value time series sequences between the key indicators, obtaining a mutual correlation coefficient between the key indicators, performing a cross-correlation comparison between the critical anomaly value time series sequences of the key indicators and non-key indicators, obtaining a mutual correlation coefficient between the key indicators and the non-key indicators, constructing a critical anomaly trend feature according to the autocorrelation coefficient of the key indicators, the mutual correlation coefficient between the key indicators, and the mutual correlation coefficient between the key indicators and the non-key indicators, and extracting a temporary compensation model for the critical anomaly acquisition period, and associating the temporary compensation model with the critical anomaly trend feature; A trend feature database is constructed, and critical abnormal trend features within several abnormal collection cycles of each node are stored in the trend feature database.

[0046] It should be further explained that, in the specific implementation process, the process of autocorrelation comparison includes: ; in, represents the autocorrelation coefficient of type index i, represents the value of type i at time t, represents the value of the standard time series of type index i at time t, and n represents the total number of time points; The process of performing a cross-correlation comparison includes: ; in, represents the mutual correlation coefficient between type index i and type index j; Represents the value of type index i at time t; represents the average value of type index i; represents the value of type index j at time t; Represents the average value of type index j.

[0047] It should be further explained that, in the specific implementation process, the process of obtaining the standard time series of various types of indicators within the historical collection period includes: A data prediction model is constructed based on deep learning. The multi-source data of each node in several historical collection cycles are used as training data. The data prediction model is trained using the training data to obtain the trained data prediction model. The multi-source data in the critical anomaly collection cycle are input into the data prediction model, and the standard time series of each type of indicator is output according to the data prediction model.

[0048] It should be further explained that, in the specific implementation process, the critical trend detection module performs a critical trend detection operation on each normal node, and according to the critical trend detection operation result, performs a parameter distribution operation on the rapid response compensation module, and the process includes: The trend feature is extracted from the numerical time series of each type of indicator in the current acquisition cycle of the normal node, and the trend feature in the current acquisition cycle of the normal node (including the autocorrelation coefficient of each type of indicator and the mutual correlation coefficient between each type of indicator) is obtained. According to the trend feature, a search and comparison is performed in the trend feature database to obtain the similarity between the trend feature and several critical abnormal trend features. If the similarity between the trend feature and the critical abnormal trend feature is greater than the preset similarity threshold, the normal node is marked as a critical warning node, and the parameters of the temporary compensation model of the critical warning node are customized according to the temporary compensation model associated with the critical abnormal trend feature. The process of screening out the top k historical fault cases with the highest cosine similarity is directly skipped, and then a rapid response compensation operation is performed on the critical warning node according to the temporary compensation model with customized parameters.

[0049] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. Adaptive data fusion and decision optimization system for heterogeneous industrial equipment, characterized by: It includes a cloud, wherein the cloud is connected to a data acquisition module, a protocol annotation module, a protocol self-consistent module, a detection decision module, a rapid response compensation module and a critical trend detection module; The data acquisition module is used to collect multi-source data of various heterogeneous industrial equipment, mark the collection time, and set the collection cycle; The protocol annotation module is used to construct an industrial connection topology map and perform initial communication protocol annotation on the industrial connection topology map; The protocol self-consistent module is used to analyze the interactive characteristics of each node in the industrial connection topology diagram and automatically match the protocol for the protocol abnormal nodes; The detection decision module is used to perform real-time detection of multi-source data of each node, and select to perform rapid response compensation operation or critical trend detection operation according to the monitoring results; The rapid response compensation module is used to generate a temporary compensation model and a joint compensation model, perform a rapid response compensation operation on the abnormal node according to the temporary compensation model, update the joint compensation model according to the rapid response compensation operation result, and then perform a joint compensation operation according to the joint compensation model; The critical trend detection module is used to extract several critical abnormal trend features of each node, perform critical trend detection operations on each normal node, and issue parameters to the rapid response compensation module according to the critical trend detection operation results.

2. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 1 is characterized in that: The protocol annotation module constructs an industrial connection topology map. The process of initial communication protocol annotation on the industrial connection topology map includes: Obtain the assembly connection relationship of several heterogeneous industrial equipment, use several heterogeneous industrial equipment as nodes of an industrial connection topology map, use the assembly connection relationship of several heterogeneous industrial equipment as the connection relationship between nodes, construct an industrial connection topology map, obtain the initial communication protocol between several heterogeneous industrial equipment, and mark the connection relationship between each node with the initial communication protocol.

3. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 2 is characterized in that: The protocol self-consistent module analyzes the interactive characteristics of each node in the industrial connection topology diagram, and the process of automatic protocol matching for protocol abnormal nodes includes: According to the initial communication protocol, the threshold intervals of various types of interaction indicators between nodes are preset, and the interaction features of the nodes with connection relationships are extracted to obtain various types of interaction indicators. If the type of interaction indicator is not within the corresponding interaction indicator threshold interval, the node with connection relationship is marked as a protocol abnormal node; The protocol fingerprint features of the protocol abnormal nodes are extracted to obtain the protocol fingerprint features, and a protocol fingerprint atlas library is constructed. According to the protocol fingerprint atlas library, an automatic protocol matching model is constructed using reinforcement learning. The protocol fingerprint features are input into the automatic protocol matching model, and the communication protocol of the protocol abnormal nodes is output. The initial communication protocol of the protocol abnormal nodes is updated according to the communication protocol.

4. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 3 is characterized in that: The process of building a fingerprint protocol graph includes: Based on the idea of ​​simulation learning, a simulator is used to generate interactive training data of different communication protocols. Baud rate analysis, data frame interval analysis and cyclic redundancy check analysis are performed on the interactive training data of different communication protocols. The baud rate probability model, data frame interval probability model and cyclic redundancy check probability model are constructed. The baud rate probability model, data frame interval probability model and cyclic redundancy check probability model are combined to construct a fingerprint protocol map.

5. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 4 is characterized in that: The detection decision module performs real-time detection on multi-source data of each node, and the process of selecting to perform rapid response compensation operation or critical trend detection operation according to the monitoring results includes: Extract the numerical time series sequence of each type of indicator in the multi-source type data of each node in the current collection cycle, preset the standard threshold interval corresponding to each type of indicator of each node, compare the numerical time series sequence of each type of indicator with the corresponding standard threshold interval, and obtain the cumulative time that each type of indicator is not in the corresponding standard threshold interval; A cumulative time threshold is preset. If the cumulative time of a node's type indicator is greater than the cumulative time threshold, a sudden abnormal alarm signal is generated and a rapid response compensation operation is performed. If the cumulative time of each type of indicator of the node is less than or equal to the cumulative time threshold, the node is marked as a normal node and a critical trend detection operation is performed.

6. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 5 is characterized in that: The process of generating a temporary compensation model by the rapid response compensation module includes: Mark the type indicator whose corresponding cumulative time is greater than the cumulative time threshold as a key indicator, mark the node to which the key indicator belongs as an abnormal node, and extract features from multi-source type data of the abnormal node in the current collection cycle based on the key indicator to obtain a feature vector set; Constructing a sudden abnormal fault library, the sudden abnormal fault library includes a number of historical fault cases, performing cosine similarity comparison between the feature data set and the number of historical fault cases, and screening out the top k historical fault cases with the highest cosine similarity; A temporary compensation model is constructed, and the compensation strategies of the top k historical fault cases are extracted to customize the parameters of the temporary compensation model. At the same time, the temporary compensation model is verified according to the multi-source type data of the abnormal nodes to obtain a verified temporary compensation model.

7. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 6 is characterized in that: The rapid response compensation module generates a joint compensation model, performs a rapid response compensation operation on abnormal nodes according to the temporary compensation model, and updates the joint compensation model according to the result of the rapid response compensation operation, including: Construct an association rule knowledge graph according to the industrial connection topology graph, wherein the association rule knowledge graph includes multi-source type data, spatiotemporal feature sequences, and joint compensation strategies between interconnected nodes, construct a joint compensation model based on deep learning, use the multi-source type data, spatiotemporal feature sequences, and joint compensation strategies as training sets and test sets, input the training sets into the joint compensation model for training until the loss function training is stable, save the model parameters, test the joint compensation model through the test set, and output the joint compensation model until it meets the preset requirements; The temporary compensation model is encapsulated and deployed, and abnormal nodes are compensated according to the temporary compensation model. The compensation time threshold is preset. When the abnormal nodes within the compensation time threshold are marked as normal nodes, the compensation parameters and effective data of the temporary compensation model are collected to update the association rule knowledge graph, and the joint compensation model is incrementally trained according to the updated association rule knowledge graph.

8. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 7 is characterized in that: The process of the rapid response compensation module performing joint compensation operations according to the joint compensation model includes: When an abnormal node is marked as a normal node, time features and spatial features are extracted from multi-source type data of the normal node and other normal nodes connected to the normal node in the current acquisition cycle to obtain a spatiotemporal feature sequence. The multi-source type data and the spatiotemporal feature sequence are input into the joint compensation model after incremental training, and a joint compensation strategy is output according to the joint compensation model.

9. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 8 is characterized in that: The process of extracting several critical abnormal trend features of each node by the critical trend detection module includes: Extract the moment when the node is marked as an abnormal node in the historical collection period, extract the numerical time series sequence and key indicators of various types of indicators between the start timestamp of the historical collection period and the moment, mark other key indicators other than the key indicators as non-key indicators, mark the numerical time series sequence as a critical abnormal numerical time series sequence, and mark the historical collection period as a critical abnormal collection period; Obtaining a standard time series sequence of key indicators within a critical anomaly acquisition period, performing an autocorrelation comparison between the critical anomaly value time series sequence of the key indicators and the standard time series sequence, obtaining an autocorrelation coefficient of the key indicators, performing a cross-correlation comparison between the critical anomaly value time series sequences between the key indicators, obtaining a mutual correlation coefficient between the key indicators, performing a cross-correlation comparison between the critical anomaly value time series sequences of the key indicators and non-key indicators, obtaining a mutual correlation coefficient between the key indicators and the non-key indicators, constructing a critical anomaly trend feature according to the autocorrelation coefficient of the key indicators, the mutual correlation coefficient between the key indicators, and the mutual correlation coefficient between the key indicators and the non-key indicators, and extracting a temporary compensation model for the critical anomaly acquisition period, and associating the temporary compensation model with the critical anomaly trend feature; A trend feature database is constructed, and critical abnormal trend features within several abnormal collection cycles of each node are stored in the trend feature database.

10. The adaptive data fusion and decision optimization system for heterogeneous industrial equipment according to claim 9 is characterized in that: The critical trend detection module performs a critical trend detection operation on each normal node, and sends parameters to the rapid response compensation module according to the critical trend detection operation result, including: Obtain the trend characteristics of the normal node in the current acquisition cycle, search and compare in the trend characteristic database according to the trend characteristics, obtain the similarity between the trend characteristics and several critical abnormal trend characteristics, if the similarity between the trend characteristics and the critical abnormal trend characteristics is greater than the preset similarity threshold, mark the normal node as a critical warning node, customize the parameters of the temporary compensation model of the critical warning node according to the temporary compensation model associated with the critical abnormal trend characteristics, and then perform a rapid response compensation operation on the critical warning node.

Images