Data sharing method and system of industrial internet

Abstract

The invention provides a data sharing method and system of an industrial internet, and relates to the technical field of the industrial internet, and the method comprises the steps: obtaining equipment data of various heterogeneous industrial equipment in the industrial internet, and carrying out the analysis of the equipment data to obtain protocol type features and structured data; then dynamically distributing the structured data to an edge computing node or a cloud computing node through a load balancing strategy according to the protocol type characteristics, the data size and the real-time level; after distribution is completed, semantic analysis is carried out on the structured data by using a deep neural network based on an attention mechanism to extract key semantic information, and the key semantic information is converted into intermediate representation data; and finally, performing differential privacy processing on the intermediate representation data according to the data sensitivity level and the query frequency, and sharing the processed data in the industrial internet. The heterogeneous data sharing efficiency of the industrial internet is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial internet, and in particular to a data sharing method and system of industrial internet. BACKGROUND

[0002] The data sharing method in the industrial internet can break down the information barriers between different devices, and is a key technology for promoting the development of digital factories. This method can realize comprehensive perception and collaborative optimization of the production process, and has broad application prospects.

[0003] The existing technical solutions process industrial equipment data by means of a deep learning model. For example, a related research proposes to use a neural network model based on an attention mechanism to directly extract semantic features from the original data stream of industrial equipment, and to share the processed data. This kind of method usually concentrates data to a unified computing node for processing.

[0004] However, these existing methods are relatively single in the data processing path, and fail to combine the real-time and resource distribution characteristics of data in the industrial scene for dynamic scheduling. At the same time, the differentiated privacy protection requirements of the processed data in sharing are not considered, resulting in a contradiction between sharing efficiency and security. Therefore, there is a technical problem of insufficient coordination between data processing efficiency and secure sharing in the prior art. SUMMARY

[0005] The present application provides a data sharing method and system of industrial internet to solve the problem of low efficiency of heterogeneous data sharing in industrial internet in the prior art.

[0006] To solve the above technical problems, in a first aspect, the present application provides a data sharing method of industrial internet, comprising:

[0007] Obtaining device data of a plurality of heterogeneous industrial equipment in the industrial internet;

[0008] Analyzing and processing the device data to obtain protocol type features and structured data;

[0009] Based on the protocol type features, and the data size and real-time level of the structured data, the structured data is dynamically distributed to edge computing nodes or cloud computing nodes through a load balancing strategy;

[0010] After the distribution is completed, the structured data is semantically analyzed by a deep neural network based on an attention mechanism to obtain key semantic information, and the key semantic information is converted into intermediate representation data;

[0011] According to the data sensitivity level and the query frequency, the intermediate representation data is processed by differential privacy, and the processed data is shared in the industrial internet.

[0012] Optionally, the structured data is semantically parsed by using the deep neural network based on the attention mechanism to obtain key semantic information, and the key semantic information is converted into intermediate representation data, including:

[0013] The structured data is grouped according to the protocol type to form a data input sequence;

[0014] The data input sequence is input into the deep neural network based on the attention mechanism, and semantic distillation is performed by the deep neural network to output key semantic information.

[0015] A semantic enhancement network is introduced to perform semantic compensation on the key semantic information, and the key semantic information after semantic compensation is converted into unified intermediate representation data according to a preset semantic mapping table.

[0016] Optionally, the data input sequence is input into the deep neural network based on the attention mechanism, and semantic distillation is performed by the deep neural network to output key semantic information, including:

[0017] The data input sequence is input into the segmentation layer of the deep neural network, and the data input sequence is segmented into multiple semantic units;

[0018] In the encoding layer of the deep neural network, the forward context feature and the backward context feature of each semantic unit are obtained by the bidirectional recurrent neural network in the encoding layer, and the forward context feature and the backward context feature are combined into complete context feature;

[0019] In the multi-head attention layer of the deep neural network, based on the complete context feature, the correlation weight distribution between different semantic units is calculated by multiple attention heads in the multi-head attention layer in parallel;

[0020] In the fusion layer of the deep neural network, the correlation weight distribution output by multiple attention heads is weighted and fused by the full connection network in the fusion layer to obtain the comprehensive correlation weight of each semantic unit.

[0021] In the selection layer of the deep neural network, the comprehensive correlation weight of each semantic unit is compared with a preset weight threshold by the comparison unit in the selection layer to filter out the semantic unit with a comprehensive correlation weight greater than the preset weight threshold as the key semantic information.

[0022] Optionally, the step of performing differential privacy processing on the intermediate representation data based on data sensitivity level and query frequency, and sharing the processed data in the Industrial Internet, includes:

[0023] Based on the content attributes of the intermediate representation data, determine the data sensitivity level corresponding to the intermediate representation data;

[0024] The query frequency is obtained by counting the number of historical queries of the intermediate data within a preset time period.

[0025] Based on the data sensitivity level and the query frequency, the target noise addition parameter is calculated using a differential privacy algorithm;

[0026] Random noise is added to the intermediate representation data based on the target noise addition parameters.

[0027] The intermediate representation data after adding random noise will be shared hierarchically on the industrial internet platform according to the access permission level of the recipient.

[0028] Optionally, the step of calculating the target noise addition parameter based on the data sensitivity level and the query frequency using a differential privacy algorithm includes:

[0029] Based on the data sensitivity level, the corresponding initial privacy parameters are obtained by querying the preset data sensitivity mapping table;

[0030] Based on the distribution range of the query frequencies, determine the frequency influence factor;

[0031] The initial privacy parameters and the frequency influence factor are input into the privacy computing engine, and intermediate privacy parameters are generated through the parameter synthesis function in the privacy computing engine.

[0032] Based on the intermediate privacy parameters, the initial noise addition parameters are calculated using the Laplace distribution function;

[0033] By combining historical performance data, the initial noise addition parameters are optimized and adjusted to obtain the target noise addition parameters.

[0034] Optionally, the step of parsing the device data to obtain protocol type characteristics and structured data includes:

[0035] Identify protocol header information from the device data, determine protocol type characteristics based on the protocol header information, and match the corresponding protocol specification template according to the protocol type characteristics;

[0036] Based on the protocol specification template, valid fields are extracted from the device data, and the valid fields are organized into structured data in a unified format.

[0037] Optionally, the step of dynamically distributing the structured data to edge computing nodes or cloud computing nodes based on the protocol type characteristics, the data volume, and the real-time requirements of the structured data through a load balancing strategy includes:

[0038] The first path weight is determined based on the protocol type characteristics, the second path weight is determined based on the data volume, and the third path weight is determined based on the real-time level.

[0039] The first path weight, the second path weight, and the third path weight are input into the path calculation model in the load balancing strategy, and the priority score of the edge computing node and the priority score of the cloud computing node are calculated through the path calculation model.

[0040] The priority scores of the edge computing nodes and the priority scores of the cloud computing nodes are compared, and the nodes with higher priority scores are identified as candidate target nodes.

[0041] Obtain the current processing load of the candidate target node; when the current processing load is less than a preset load threshold, distribute the structured data to the candidate target node.

[0042] When the current processing load of the candidate target node is greater than the preset load threshold, the node with the second highest priority score is selected as the actual target node, and the structured data is distributed to the actual target node.

[0043] After distribution is complete, update the load status of the edge computing node and the cloud computing node.

[0044] Secondly, this application provides an industrial internet data sharing system, comprising:

[0045] The acquisition module is used to acquire device data from various heterogeneous industrial devices in the Industrial Internet.

[0046] The parsing module is used to parse and process the device data to obtain protocol type characteristics and structured data;

[0047] The distribution module is used to dynamically distribute the structured data to edge computing nodes or cloud computing nodes based on the characteristics of the protocol type, the data volume and real-time level of the structured data, and through a load balancing strategy.

[0048] The conversion module is used to perform semantic parsing on the structured data using a deep neural network based on an attention mechanism after the distribution is completed, to obtain key semantic information, and to convert the key semantic information into intermediate representation data.

[0049] The processing module is used to perform differential privacy processing on the intermediate representation data according to the data sensitivity level and query frequency, and to share the processed data in the Industrial Internet.

[0050] Thirdly, this application provides an electronic device, comprising:

[0051] Memory, used to store computer programs;

[0052] A processor, used to execute the computer program to implement the steps of the data sharing method for the Industrial Internet as described in the first aspect above.

[0053] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps of the industrial internet data sharing method described in the first aspect above.

[0054] The data sharing method for the Industrial Internet provided in this application has the following beneficial effects:

[0055] First, by acquiring equipment data from various heterogeneous industrial devices and parsing it to obtain protocol type characteristics and structured data, it is possible to accurately identify and integrate data from different sources, providing a unified foundation for subsequent processing. Then, based on protocol type, data volume, and real-time level, data is dynamically distributed to edge or cloud computing nodes, thereby flexibly allocating computing resources to meet the processing speed and computing power requirements of different data.

[0056] Next, semantic parsing is performed using a deep neural network based on an attention mechanism and the data is converted into intermediate representation data. This can effectively extract the core semantics of data from different protocols and enable semantic interoperability between heterogeneous data. Finally, after differential privacy processing based on data sensitivity and query frequency, the data is shared, which can support the secure flow and use of data while protecting data privacy.

[0057] Furthermore, the structured data is first grouped into an input sequence according to the protocol type, and then input into a deep neural network based on the attention mechanism for semantic distillation to extract key semantic information. Next, a semantic enhancement network is introduced to perform semantic compensation on this information. Finally, it is converted into unified intermediate representation data according to a preset mapping table. This process can improve the accuracy and completeness of cross-protocol semantic parsing and ensure that core information is preserved when heterogeneous data is uniformly expressed.

[0058] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 A flowchart illustrating a data sharing method for the Industrial Internet provided in this application embodiment;

[0061] Figure 2 This application provides a schematic diagram illustrating a specific implementation of a data sharing method for the Industrial Internet.

[0062] Figure 3 This is a schematic diagram of the structure of an industrial internet data sharing system provided in an embodiment of this application. Detailed Implementation

[0063] In the field of industrial internet data sharing, existing technologies typically rely on centralized deep learning models to perform semantic processing and sharing of equipment data. This approach employs a single data computation path in its processing flow, lacking targeted scheduling arrangements to address the real-time requirements and resource distribution characteristics of different data in industrial scenarios. Furthermore, these methods have limited ability to balance the need for differentiated data protection with sharing efficiency during the data sharing process, making it difficult to achieve effective coordination between efficiency and security.

[0064] To address the aforementioned issues, this application provides a data sharing method for the Industrial Internet. The core idea of ​​this method is as follows: First, acquire and parse heterogeneous device data to extract protocol features and structured data. Then, based on the protocol type, data volume, and real-time requirements, dynamically allocate data to edge or cloud computing nodes for processing through a load balancing strategy. Next, use a deep neural network based on an attention mechanism to perform semantic parsing on the data and convert it into a unified intermediate representation. Finally, perform differential privacy processing on the intermediate representation based on data sensitivity and query frequency before sharing.

[0065] Therefore, this method adapts to the processing needs of different data through dynamic load balancing, improves the semantic interoperability of heterogeneous data through semantic parsing, and balances efficiency and security in the sharing process by combining differentiated privacy protection, thereby effectively solving the problem of insufficient data processing efficiency and secure sharing synergy.

[0066] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0067] Example 1

[0068] The core of this application is to provide a data sharing method for the Industrial Internet, and a flowchart of one specific implementation is shown below. Figure 1 As shown, the method includes:

[0069] Step 101: Obtain device data of various heterogeneous industrial devices in the Industrial Internet.

[0070] In step 101, the Industrial Internet refers to a network that connects various types of equipment and systems in industrial production, and heterogeneous industrial equipment refers to industrial control equipment and sensing equipment with different brands or communication protocols.

[0071] Device data refers to the raw data generated by various devices in the Industrial Internet during operation. This data includes device operating status readings, operation command records, physical quantity measurements collected by sensors, and device identification information, and is usually transmitted in the specific communication protocol format followed by the device itself.

[0072] Step 102: Parse the device data to obtain protocol type characteristics and structured data.

[0073] Among them, the protocol type feature is used to characterize the type of communication protocol followed by the device data, and structured data refers to data records that are organized according to a predetermined format and contain explicit field names and field values.

[0074] In this embodiment, step 102 includes the following process:

[0075] Step 1021: Identify protocol header information from the device data, determine protocol type characteristics based on the protocol header information, and match the corresponding protocol specification template according to the protocol type characteristics.

[0076] In step 1021, the protocol header information refers to a specific byte sequence at the start position of the device data frame, which is used to identify the protocol type; the protocol specification template is a predefined set of rules that describes the structure of a specific protocol data frame, which includes the definition of field position, field length and field meaning.

[0077] In this embodiment of the application, a byte sequence of a predetermined length is first read from the starting position of the device data as protocol header information. Then, the protocol header information is compared with a pre-stored protocol feature library to determine the protocol type feature. Finally, the protocol template library is queried according to the determined protocol type feature to obtain the corresponding protocol specification template.

[0078] In practical applications, taking a raw data hexadecimal byte stream 010300000002C40B from a control device of brand A as an example, the first two bytes 0103 are read as protocol header information. After comparison with the feature library, it is determined that the protocol header information corresponds to the Modbus RTU protocol, thus obtaining the protocol type feature as Modbus RTU. Subsequently, based on the protocol type feature, the corresponding Modbus RTU protocol specification template is matched from the template library. This template defines the byte position and length of the device address, function code, data field, and checksum.

[0079] Step 1022: Based on the protocol specification template, extract valid fields from the device data and organize the valid fields into structured data in a unified format.

[0080] In step 1022, a valid field refers to a data unit with actual semantics that is parsed from the device data according to the protocol specification template; a structured data in a unified format refers to a standardized data record formed by organizing all valid fields according to fixed field names, field values ​​and timestamp formats.

[0081] In this embodiment of the application, based on the field position and length defined in the matched protocol specification template, the corresponding byte sequence is extracted from the device data and parsed into valid fields. Then, all the parsed valid fields, together with the data collection timestamp, are filled into a preset unified data format to generate structured data.

[0082] In practical applications, continuing the above example, based on the matched Modbus RTU protocol specification template, valid fields are extracted from byte stream 010300000002C40B. Byte 1 (01) is parsed as the device address field, byte 2 (03) as the function code field, bytes 3 to 6 (00000002) as the data field, and bytes 7 to 8 (C40B) as the check field. Finally, these field values, along with the collection timestamp of May 27, 2024, 10:30:25:345, are organized into a structured data record containing field names, field values, and a timestamp.

[0083] This application can accurately identify the communication protocols of different industrial equipment through step 102, and convert the raw data into structured data with a unified format and clear semantics, laying the foundation for subsequent data processing and sharing.

[0084] Step 103: Based on the characteristics of the protocol type, the data volume and real-time level of the structured data, the structured data is dynamically distributed to edge computing nodes or cloud computing nodes through a load balancing strategy.

[0085] The network consisting of edge computing nodes and cloud computing nodes is a component of the industrial internet implementation architecture. Edge computing nodes are independent computing units deployed near the equipment, responsible for data processing and computing tasks, while heterogeneous industrial equipment are physical devices that directly participate in production control. High real-time status control data and small-to-medium volume real-time monitoring data are distributed to edge computing nodes, while non-real-time large-volume historical records and complex analysis task data are distributed to cloud computing nodes.

[0086] In this embodiment, step 103 includes the following process:

[0087] Step 1031: Determine the first path weight based on the protocol type characteristics, determine the second path weight based on the data volume, and determine the third path weight based on the real-time level.

[0088] In step 1031, the first path weight is a value that quantifies the degree of influence of protocol type characteristics on the selection of computing nodes, the second path weight is a value that quantifies the degree of influence of data volume on the selection of computing nodes, and the third path weight is a value that quantifies the degree of influence of real-time level on the selection of computing nodes.

[0089] In this embodiment, the first path weight corresponding to the protocol type feature is first queried according to the pre-established protocol type weight table. Then, the second path weight corresponding to the data size is calculated according to the data volume weight function. Finally, the third path weight corresponding to the real-time level is queried according to the real-time weight table.

[0090] In practical applications, for the aforementioned structured data, its protocol type characteristic is Modbus RTU. Looking up the protocol type weight table, the first path weight is 0.2. The data size of this structured data is 8 bytes. The second path weight is calculated using a data size weight function. Specifically, the weight is 0.3 when the data size is less than 100 bytes, 0.5 when the data size is between 100 and 1000 bytes, and 0.7 when the data size is greater than 1000 bytes. Therefore, the second path weight corresponding to 8 bytes of data is 0.3. The real-time level of this structured data is high. Looking up the real-time weight table, the third path weight corresponding to high level is 0.5. It should be noted that the specific expression used for the data size weight function in this embodiment is not specifically limited; it can be set according to the actual situation.

[0091] Step 1032: Input the first path weight, the second path weight and the third path weight into the path calculation model in the load balancing strategy, and calculate the priority score of the edge computing node and the priority score of the cloud computing node through the path calculation model.

[0092] In step 1032, the path calculation model is a mathematical model used to integrate multiple weights and calculate the priority of each computing node. The path calculation model can be implemented in various ways, including: a weighted summation model, a rule-based decision tree, or a lightweight multilayer perceptron. This embodiment does not specifically limit the specific structure of the model and can set it accordingly according to the actual situation. The priority score is a quantitative value that represents the priority of distributing data to a certain computing node.

[0093] In this embodiment, the obtained first path weight, second path weight, and third path weight are used as inputs. The priority scores of the edge computing nodes and the cloud computing nodes are calculated respectively using a weighted summation formula in the path calculation model. For example, the path calculation model is a weighted summation model, wherein the formula for calculating the priority score of the edge computing nodes is as follows: The formula for calculating the priority score of a cloud computing node is as follows: ,in and These are priority scores for edge computing nodes and cloud computing nodes, respectively. , , These represent the first path weight, the second path weight, and the third path weight, respectively. and These are the allocation coefficients for the first path weight on edge computing nodes and cloud computing nodes, respectively. and These are the allocation coefficients for the second path weights on edge computing nodes and cloud computing nodes, respectively. and These are the allocation coefficients of the third path weight on edge computing nodes and cloud computing nodes, respectively. These coefficients are obtained through regression analysis of historical task execution logs, and the optimization objective is to minimize the overall task processing latency.

[0094] In practical applications, the allocation coefficient is set as follows: =0.8, =0.2, =0.3, =0.7, =0.9, =0.1 (all are system preset values, which will not be elaborated further); Substituting the weights of the first path (0.2), the second path (0.3), and the third path (0.5) into the formula, the priority score of the edge computing node is obtained. =0.2×0.8+0.3×0.3+0.5×0.9=0.16+0.09+0.45=0.70, the priority score of the cloud computing node. =0.2×0.2+0.3×0.7+0.5×0.1=0.04+0.21+0.05=0.30.

[0095] In other examples, for structured data from an emergency stop button on a production line, the content represents an "emergency stop" instruction. According to business rules, the real-time level of such safety control instructions is defined as "highest." The data volume is very small, only a few bytes. According to the pre-configured strategy, data with the highest real-time level must respond within milliseconds. Therefore, edge computing nodes receive extremely high path weights due to their proximity. The extremely small data volume has negligible impact on network bandwidth and node computing resource consumption. Therefore, after considering these characteristics, the path calculation model calculates that the priority score of edge computing nodes is much higher than that of cloud computing nodes.

[0096] Step 1033: Compare the priority scores of the edge computing nodes with the priority scores of the cloud computing nodes, and determine the node with the higher priority score as the candidate target node.

[0097] In practical applications, the calculated priority score of the edge computing node (0.70) is compared with that of the cloud computing node (0.30). Since 0.70 is greater than 0.30, the edge computing node is determined as a candidate target node.

[0098] Step 1034: Obtain the current processing load of the candidate target node. When the current processing load is less than a preset load threshold, distribute the structured data to the candidate target node.

[0099] In step 1034, the current processing load refers to the percentage of utilization of the computing resources of the computing node that are being used to execute tasks at the moment the distribution instruction is received. This percentage reflects the current busyness of the node.

[0100] The load threshold is a pre-set critical value used to determine whether a computing node is in an overloaded state. In this embodiment of the application, the value of the preset load threshold is not specifically limited, but can be set according to the actual situation.

[0101] In this embodiment of the application, a status query request is sent to the candidate target node to obtain its current processing load, which is usually expressed as a percentage of the node's central processing unit utilization. Then, the obtained current processing load is compared with a preset load threshold. If the current processing load is less than the load threshold, the structured data is sent to the candidate target node.

[0102] In practical applications, the CPU utilization rate of the candidate target node determined by the query, i.e. the edge computing node, is 75%, and the preset load threshold is 80%. Since 75% is less than 80%, the structured data is distributed to the edge computing node.

[0103] Step 1035: When the current processing load of the candidate target node is greater than the preset load threshold, select the node with the second highest priority score as the actual target node, and distribute the structured data to the actual target node.

[0104] In practical applications, assuming that the CPU utilization rate of the edge computing node is 85%, which is greater than the preset load threshold of 80%, the cloud computing node with the second highest priority score is selected as the actual target node, and the structured data is distributed to the cloud computing node.

[0105] Step 1036: After distribution is complete, update the load status of the edge computing node and the cloud computing node.

[0106] In this embodiment of the application, after the data distribution operation is completed, the load status records of the edge computing node and the cloud computing node are updated according to the distribution target node and the estimated resource consumption of the node in processing the new data.

[0107] In practical applications, after the distribution to the edge computing nodes is completed, it is estimated based on empirical values ​​that this data processing will increase the utilization of their central processing units by 2 percentage points. Therefore, the load status of the edge computing nodes is updated from 75% to 77%, while the load status of the cloud computing nodes remains unchanged.

[0108] This application, through step 103, can intelligently select the most suitable computing node based on multiple characteristics of the data, and automatically switch when the node load is too high, thereby realizing dynamic optimization of computing resources and improving data processing efficiency and overall system stability.

[0109] Step 104: After the distribution is completed, the structured data is semantically parsed using a deep neural network based on an attention mechanism to obtain key semantic information, and the key semantic information is converted into intermediate representation data.

[0110] Among them, key semantic information refers to data units extracted from structured data that can characterize the core state or key parameters of the device. The intermediate representation data is a unified standardized data format that is independent of the original protocol and is used to carry key semantic information.

[0111] Furthermore, the network structure of the deep neural network based on the attention mechanism can be as described in A1 to A5 below, and will not be repeated here. However, this embodiment does not specifically limit the structural design of the network, and the model was trained before the application stage. For example:

[0112] Historical industrial equipment data streams are collected, and domain experts annotate data segments containing key information with key semantic information labels based on business knowledge. Simultaneously, a semantic knowledge base is constructed, including alarm codes, parameter descriptions, and standard units from equipment manuals. Then, the grouped structured data sequence is used as input, and the annotated key semantic information sequence is used as the supervision target. The network employs a cross-entropy loss function and uses the Adam optimizer for end-to-end training. The training objective is to minimize the difference between the predicted key semantic information and the true labels.

[0113] In this embodiment, step 104 includes the following process, such as... Figure 2 As shown:

[0114] Step 1041: Group the structured data according to protocol type to form a data input sequence.

[0115] In step 1041, the data input sequence is a set of multiple structured data of the same protocol type arranged in chronological or logical order.

[0116] In this embodiment of the application, the structured data that arrives at the computing node after being distributed in step 103 is first classified according to its protocol type characteristics. The structured data with the same protocol type characteristics are collected together and arranged in timestamp order to form a data input sequence for each protocol type.

[0117] In practical applications, assuming that the structured data received by the edge computing node contains 10 data entries of the Modbus RTU protocol type and 5 data entries of the Profinet protocol type, the 10 Modbus RTU data entries are arranged from earliest to latest according to their timestamps to form the Modbus RTU data input sequence. At the same time, the 5 Profinet data entries are arranged in the same way to form the Profinet data input sequence.

[0118] Step 1042: Input the data input sequence into a deep neural network based on an attention mechanism, and perform semantic distillation through the deep neural network to output key semantic information. Semantic distillation describes the process of extracting key information from data using a deep neural network. The process involves semantic distillation through the deep neural network (i.e., encoding the input sequence through a series of neural network layers, focusing on key information, and filtering out core semantics) to output key semantic information. Further details are omitted here.

[0119] Step 1042 may specifically include the following steps:

[0120] A1: Input the data input sequence into the segmentation layer of the deep neural network to perform semantic segmentation on the data input sequence, dividing it into multiple semantic units.

[0121] In step A1, a semantic unit is the smallest data segment in the data input sequence that has independent semantics.

[0122] In the embodiments of this application, the segmentation layer of the deep neural network divides each piece of structured data in the data input sequence into multiple smaller, semantically independent segments based on field separators and predefined semantic boundary rules in the data. These segments are semantic units.

[0123] In practical applications, taking a Modbus RTU structured data "Device address: 01, Function code: 03, Data: 00000002" as an example, the segmentation layer divides it into three semantic units based on delimiters such as colons and commas, namely "Device address: 01", "Function code: 03" and "Data: 00000002".

[0124] In another practical application, a structured data message from a CNC machine tool, with a custom text protocol, is: "ALARM:123; AXIS:X; CURRENT:15.6A; TIME:20240527103025". The segmentation layer divides this data into four semantic units based on the semicolon separator: "ALARM:123", "AXIS:X", "CURRENT:15.6A", and "TIME:20240527103025".

[0125] A2: In the encoding layer of the deep neural network, the forward context features and backward context features of each semantic unit are obtained through the bidirectional recurrent neural network in the encoding layer, and the forward context features and backward context features are combined into complete context features.

[0126] In step A2, the forward context feature refers to the encoding of context information extracted from the beginning of the sequence to the current semantic unit, and the backward context feature refers to the encoding of context information extracted from the end of the sequence to the current semantic unit. The complete context feature is a comprehensive feature vector formed by concatenating and combining the forward and backward features.

[0127] In this embodiment, the bidirectional recurrent neural network in the encoding layer first processes the semantic unit sequence in the forward direction, generating a forward context feature vector for each semantic unit. Then, it processes the sequence again in the reverse direction, generating a backward context feature vector for each semantic unit. Finally, the forward and backward feature vectors corresponding to the same semantic unit are concatenated along the feature dimension to form the complete context feature vector of the semantic unit.

[0128] In practical applications, for the semantic unit "data: 00000002", the bidirectional recurrent neural network processes the data in forward order to obtain its forward context feature vector [0.1, 0.3, 0.5], processes it in reverse order to obtain its backward context feature vector [0.2, 0.4, 0.6], and finally concatenates these two vectors to obtain its complete context feature vector [0.1, 0.3, 0.5, 0.2, 0.4, 0.6].

[0129] A3: In the multi-head attention layer of the deep neural network, based on the complete context features, the association weight distribution between different semantic units is calculated in parallel by multiple attention heads in the multi-head attention layer.

[0130] In step A3, the association weight distribution is used to quantify the importance of one semantic unit to understanding another semantic unit.

[0131] In this embodiment, the multi-head attention layer receives the complete context feature vectors of all semantic units as input. Multiple independent attention heads work simultaneously within it. Each attention head outputs a set of pairwise association weight matrices between semantic units by calculating the dot product attention between the query vector, key vector, and value vector.

[0132] In practical applications, two attention heads are set up for the three semantic units "Device Address: 01", "Function Code: 03", and "Data: 00000002". The weight matrix calculated by the first attention head may indicate that the association weight between "Function Code" and "Data" is 0.8, and the weight matrix calculated by the second attention head may indicate that the association weight between "Device Address" and "Function Code" is 0.6.

[0133] A4: In the fusion layer of the deep neural network, the association weight distributions of multiple attention heads are weighted and fused through the fully connected network in the fusion layer to obtain the comprehensive association weight of each semantic unit.

[0134] In step A4, the comprehensive association weight is the final quantitative score of the overall importance of a single semantic unit after integrating multiple attention perspectives.

[0135] In this embodiment, the fully connected network in the fusion layer receives the association weight matrix output by all attention heads. First, it concatenates the weights of each semantic unit under different attention heads, and then performs nonlinear transformation and weighted summation through the fully connected network. Finally, it outputs a scalar value as its comprehensive association weight for each semantic unit.

[0136] In practical applications, for the semantic unit "data: 00000002", its weight in the first attention head is 0.8, and its weight in the second attention head is 0.1. The fully connected network transforms the input vector [0.8, 0.1], for example, through the formula... Calculate the overall association weight, where This represents the overall association weight of semantic units. and These are the coefficients learned by the fully connected network. and It is the input weight, set , The overall association weight is .

[0137] A5: In the selection layer of the deep neural network, the comparison unit in the selection layer compares the comprehensive association weight of each semantic unit with a preset weight threshold to filter out semantic units whose comprehensive association weight is greater than the preset weight threshold as key semantic information.

[0138] In this embodiment of the application, the comparison unit in the selection layer compares the comprehensive association weight calculated for each semantic unit with a preset weight threshold, and then filters out all semantic units whose comprehensive association weight is greater than the preset weight threshold. These finally filtered semantic units are determined to be key semantic information.

[0139] In practical applications, a preset weight threshold of 0.5 is set. The comprehensive association weight of the semantic unit "Device Address: 01" is 0.2, the comprehensive association weight of the semantic unit "Function Code: 03" is 0.45, and the comprehensive association weight of the semantic unit "Data: 00000002" is 0.59. After comparison, the comprehensive association weight of 0.59 is greater than 0.5. Therefore, the semantic unit "Data: 00000002" is selected as the key semantic information.

[0140] In another practical application, a preset weight threshold of 0.6 was set. After calculation, the comprehensive association weight of the semantic unit "ALARM:123" was 0.85, "CURRENT:15.6A" was 0.70, and the weights of the remaining units were all below 0.6. Therefore, "ALARM:123" and "CURRENT:15.6A" were selected as key semantic information.

[0141] Step 1043: Introduce a semantic enhancement network to perform semantic compensation on the key semantic information, and convert the semantically compensated key semantic information into unified intermediate representation data according to a preset semantic mapping table.

[0142] In step 1043, the semantic enhancement network is an auxiliary neural network used to supplement and enrich potentially missing context or attributes in key semantic information. It consists of a semantic knowledge base and a feedforward neural network. The semantic knowledge base stores cross-protocol semantic association rules, and the feedforward neural network is responsible for querying the knowledge base based on the input key semantic information and generating compensation content. The semantic mapping table is a lookup table that defines the conversion rules between the original key semantic information and the standardized intermediate representation fields. It should be noted that this embodiment does not specifically limit the structure of the semantic enhancement network, and corresponding settings can be made according to the actual situation.

[0143] During the model training phase, for example, the semantic augmentation network can retrieve the corresponding text description from the semantic knowledge base based on the identifier in the key semantic information, such as the alarm code "123". The training objective is to learn to stably map the original identifier and the retrieval description to the preset standardized field names. The training data consists of triples: the original identifier, the retrieval description, and the standardized field names.

[0144] In this embodiment of the application, the selected key semantic information is input into a pre-trained semantic enhancement network. Based on the cross-protocol semantic knowledge stored in its internal storage, the network adds supplementary information such as context description or physical units to the key semantic information. Then, the complete key semantic information after semantic compensation is queried in a preset semantic mapping table to find the corresponding standardized field names and formats, and finally generates unified intermediate representation data.

[0145] In practical applications, the key semantic information "data: 00000002" is transformed into "pressure reading: 00000002, unit: megapascal" after being compensated by a semantic enhancement network. Then, the semantic mapping table is queried, which specifies that "pressure reading" is mapped to "Pressure" and "megapascal" is mapped to "MPa". The numerical values ​​remain unchanged, so the intermediate representation data "Pressure: 00000002, Unit: MPa" is obtained.

[0146] In another practical application, key semantic information is input into a semantic enhancement network. This network has a pre-built industrial alarm code library and a parameter unit mapping table. Through querying, "ALARM: 123" is compensated to "Alarm Type: X-axis Overload", and "CURRENT: 15.6A" is compensated to "Current Value: 15.6, Unit: Ampere". Subsequently, according to the semantic mapping table, "Alarm Type" is mapped to "AlarmType", "Current Value" is mapped to "Current", and "Ampere" is mapped to "A". The final intermediate representation data is: {"AlarmType": "X-axis Overload", "Current": 15.6, "Unit": "A"}.

[0147] This application can accurately extract core semantic information from massive heterogeneous device data through step 104, and convert it into a format that machines can uniformly understand and process through semantic enhancement and standardized mapping, providing a solid data foundation for cross-platform and cross-protocol data fusion and advanced applications.

[0148] Step 105: Perform differential privacy processing on the intermediate representation data according to the data sensitivity level and query frequency, and share the processed data in the Industrial Internet.

[0149] Among them, the data sensitivity level is an indicator that quantifies and classifies the confidentiality of the information contained in the intermediate representation data, and the query frequency is a statistical value of the number of times the intermediate representation data is accessed within a specific time period.

[0150] In this embodiment, step 105 includes the following process:

[0151] Step 1051: Determine the data sensitivity level corresponding to the intermediate representation data based on the content attributes of the intermediate representation data.

[0152] In step 1051, the content attribute refers to the specific business meaning represented by the intermediate data, such as equipment pressure, temperature or operating status; the data sensitivity level refers to multiple confidentiality levels that are pre-divided based on the degree of impact of the data content on industrial production or the severity of the consequences that may be caused by leakage, such as public, internal, secret, etc.

[0153] In this embodiment of the application, the business meaning represented by the field names and values ​​of the intermediate data is first analyzed, and then the business meaning is mapped to the corresponding data sensitivity level according to the pre-set hierarchical rules.

[0154] In practical applications, for the intermediate data “Pressure: 00000002, Unit: MPa”, its content attribute is “equipment pressure reading”. Then, according to the classification rule “equipment core process parameters are of high sensitivity level”, its data sensitivity level is determined to be of high sensitivity level.

[0155] Step 1052: Count the number of historical queries of the intermediate representation data within a preset time period to obtain the query frequency.

[0156] In step 1052, the preset time period refers to a time length set in advance for statistical query frequency, such as the most recent 24 hours, the most recent week, or the most recent month; the number of historical queries refers to the total number of times that a certain data in the past time period was accessed and read by the data requester; the query frequency refers to the number of historical queries that the data in the preset time period was accessed, which reflects the popularity or attention of the data.

[0157] In this embodiment of the application, the intermediate data is retrieved from the access log of the data sharing platform. All queried records of the intermediate data within the most recent preset time period are retrieved, and the total number of records is counted. This total number of records is the query frequency.

[0158] In practical applications, the preset time period is 24 hours. Querying the access logs of the intermediate data "Pressure: 00000002, Unit: MPa" in the past 24 hours, a total of 150 query records were found, so its query frequency is 150 times.

[0159] Step 1053: Based on the data sensitivity level and the query frequency, calculate the target noise addition parameter using a differential privacy algorithm.

[0160] Among them, the target noise addition parameter is a specific value that is optimized and adjusted to control the intensity of random noise added to the data. This parameter directly determines the level of privacy protection.

[0161] Step 1053 may specifically include the following steps:

[0162] B1: Based on the data sensitivity level, the corresponding initial privacy parameters are obtained by querying the preset data sensitivity mapping table.

[0163] In step B1, the data sensitivity mapping table is a lookup table that defines the correspondence between data sensitivity levels and initial privacy parameters. The initial privacy parameters are preliminary privacy protection strength values ​​used to calculate noise, obtained directly from the preset mapping table based on the data sensitivity level.

[0164] In this embodiment, the privacy computing engine queries its internally stored data sensitivity mapping table based on the input data sensitivity level to directly retrieve the corresponding initial privacy parameter value. Furthermore, in the differential privacy domain, the aforementioned privacy protection strength is controlled by privacy parameters; the smaller the privacy control parameter, the stronger the privacy protection.

[0165] In practical applications, the data sensitivity mapping table specifies that the initial privacy parameter is 0.1 for high sensitivity level, 0.5 for medium sensitivity level, and 1.0 for low sensitivity level; then for high sensitivity level data, the initial privacy parameter is found to be 0.1.

[0166] B2: Determine the frequency influence factor based on the distribution range of the query frequency.

[0167] In step B2, the distribution range is obtained by counting the number of queries of all intermediate representation data within a preset time period and analyzing the maximum, minimum and concentration intervals of these numbers. It is used to quantify the differences in the frequency of access to different data. The frequency impact factor is a correction coefficient used to adjust the privacy protection strength according to the query frequency. Its value decreases as the query frequency increases.

[0168] In other words, the smaller the frequency impact factor, the more frequently the data is queried, and the correspondence between the frequency impact factor and the frequency range can be calibrated by analyzing the historical correlation between data popularity and privacy risks.

[0169] In this embodiment of the application, the privacy computing engine compares the query frequency with several preset frequency intervals to determine the distribution range to which it belongs, and then determines the corresponding frequency influence factor based on the distribution range.

[0170] In practical applications, the preset frequency ranges are low frequency 0-50 times, medium frequency 51-200 times, and high frequency above 200 times; among them, the query frequency of 150 times belongs to the medium frequency range, and the frequency influence factor corresponding to this range is 0.8.

[0171] B3: Input the initial privacy parameters and the frequency influence factor into the privacy computing engine, and generate intermediate privacy parameters through the parameter synthesis function in the privacy computing engine.

[0172] In step B3, the parameter synthesis function is a mathematical function used to fuse the initial privacy parameters and the frequency influence factor to generate intermediate privacy parameters, defining how to fuse the initial privacy parameters and the frequency influence factor.

[0173] In one embodiment, the privacy computing engine substitutes the received initial privacy parameters and frequency influence factor into its internal parameter synthesis function for calculation. This function is typically a multiplication function, meaning that the intermediate privacy parameter is equal to the initial privacy parameter multiplied by the frequency influence factor.

[0174] In practical applications, the initial privacy parameter is 0.1, the frequency influence factor is 0.8, and then the intermediate privacy parameter is calculated by multiplication function: 0.1 × 0.8 = 0.08.

[0175] B4: Based on the intermediate privacy parameters, calculate the initial noise addition parameters using the Laplace distribution function.

[0176] In step B4, the Laplace distribution function is a probability distribution function used to generate noise that meets the differential privacy requirements. Its scale parameter is the noise addition parameter. It should be noted that this embodiment does not specifically limit the specific expression of the Laplace distribution function, and can be set accordingly according to the actual situation.

[0177] In this embodiment, the privacy computing engine uses the intermediate privacy parameters as the privacy budget for differential privacy, and calculates the initial noise addition parameters based on the global sensitivity of the query function and the scaling parameter calculation formula of the Laplace distribution function.

[0178] In practical applications, the global sensitivity of the query function is set to 1.0, and the formula for calculating the Laplace distribution scaling parameter b is as follows: ,in The scale parameter of the Laplace distribution is represented. This refers to global sensitivity, such as 1.0. This is an intermediate privacy parameter, such as 0.08, which is substituted into the formula to calculate the initial noise addition parameter. .

[0179] B5: Based on historical performance data, the initial noise addition parameters are optimized and adjusted to obtain the target noise addition parameters.

[0180] In step B5, historical effect data is log data that records the evaluation results of data availability and privacy protection after using different noise parameters in the past. Specifically, historical effect data includes records of data usage after adding noise in the past, as well as quantitative evaluation results of these records in terms of data availability and privacy protection strength.

[0181] In this embodiment, the privacy computing engine searches the historical effect database to find the effect scores corresponding to different noise parameters in scenarios similar to the current data. Then, it fine-tunes the initial noise addition parameters by interpolation or finding the optimal value, and finally obtains the optimized target noise addition parameters.

[0182] In practical applications, historical data show that when the noise parameter is between 10.0 and 15.0, the data availability score meets the standard. The initial noise addition parameter is 12.5, which is within this range, so it is determined to be the target noise addition parameter of 12.5.

[0183] Step 1054: Add random noise to the intermediate representation data according to the target noise addition parameters.

[0184] In this embodiment of the application, the target noise addition parameter is used as the scale parameter of the Laplace distribution to generate a random noise value that satisfies the distribution. Then, the random noise value is added to the numerical field in the intermediate representation data to obtain the data after noise addition.

[0185] In practical applications, the target noise addition parameter is 12.5. A random noise value of +3.2 is generated according to the Laplace distribution. The original intermediate representation data has a numerical field of 00000002, which is decimal 2. After adding noise, it becomes 2 + 3.2 = 5.2.

[0186] Step 1055: The intermediate representation data after adding random noise is shared hierarchically on the industrial internet platform according to the access permission level of the recipient.

[0187] In step 1055, the recipient's access permission level refers to the different data access scopes and control permissions set for different data requesters according to the data sharing policy.

[0188] In this embodiment, based on the access permission level of the data requester, it is decided to share all or part of the fields of the intermediate representation data after noise addition with the other party. The data granularity obtained by the receivers with different access permission levels is different, thereby realizing data sharing under the premise of security.

[0189] In practical applications, receiver A has a high access level and can obtain the complete data "Pressure: 5.2, Unit: MPa" after noise has been added; receiver B has a low access level and can only obtain the anonymized data "PressureRange: Normal".

[0190] This application, through step 105, can dynamically adjust the privacy protection strength based on the intrinsic value of the data and the access frequency, and share the protected data with different objects as needed, thereby effectively promoting data circulation in the industrial internet while ensuring data security.

[0191] Example 2

[0192] Figure 3 A schematic diagram of the structure of an industrial internet data sharing system provided in this application embodiment is shown below. Figure 3 As shown, the detailed implementation section describes:

[0193] The acquisition module 31 is used to acquire equipment data of various heterogeneous industrial devices in the Industrial Internet.

[0194] The parsing module 32 is used to parse and process the device data to obtain protocol type features and structured data.

[0195] The distribution module 33 is used to dynamically distribute the structured data to edge computing nodes or cloud computing nodes based on the characteristics of the protocol type, the data volume and real-time level of the structured data, and through a load balancing strategy.

[0196] The conversion module 34 is used to perform semantic parsing on the structured data using a deep neural network based on an attention mechanism after the distribution is completed, to obtain key semantic information, and to convert the key semantic information into intermediate representation data.

[0197] The processing module 35 is used to perform differential privacy processing on the intermediate representation data according to the data sensitivity level and query frequency, and to share the processed data in the Industrial Internet.

[0198] The industrial internet data sharing system of this application embodiment is used to implement the aforementioned industrial internet data sharing method. Therefore, the specific implementation of the industrial internet data sharing system can be found in the embodiment section of the industrial internet data sharing method above. The specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0199] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described industrial internet data sharing methods.

[0200] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described industrial internet data sharing methods.

[0201] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as USB flash drives, read-only memory, random access memory, portable hard drives, magnetic disks, or optical disks.

[0202] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the data sharing method for the Industrial Internet.

[0203] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0204] The above provides a detailed description of the data sharing method and system for the Industrial Internet provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A data sharing method for the Industrial Internet, characterized in that, include: Acquire device data from various heterogeneous industrial devices in the Industrial Internet; The device data is parsed to obtain protocol type characteristics and structured data; Based on the characteristics of the protocol type, and the data volume and real-time level of the structured data, the structured data is dynamically distributed to edge computing nodes or cloud computing nodes through a load balancing strategy. After distribution is completed, the structured data is semantically parsed using a deep neural network based on an attention mechanism to obtain key semantic information, and the key semantic information is then converted into intermediate representation data. Differential privacy processing is performed on the intermediate representation data based on the data sensitivity level and query frequency, and the processed data is shared in the Industrial Internet.

2. The method according to claim 1, characterized in that, The step of using a deep neural network based on an attention mechanism to perform semantic parsing on the structured data to obtain key semantic information, and converting the key semantic information into intermediate representation data, includes: The structured data is grouped according to protocol type to form a data input sequence; The data input sequence is fed into a deep neural network based on an attention mechanism, and semantic distillation is performed by the deep neural network to output key semantic information. A semantic enhancement network is introduced to perform semantic compensation on the key semantic information, and the semantically compensated key semantic information is converted into unified intermediate representation data according to a preset semantic mapping table.

3. The method according to claim 2, characterized in that, The step of inputting the data input sequence into a deep neural network based on an attention mechanism, performing semantic distillation through the deep neural network, and outputting key semantic information includes: The data input sequence is input into the segmentation layer of a deep neural network to perform semantic segmentation on the data input sequence, dividing it into multiple semantic units; In the encoding layer of the deep neural network, the forward context features and backward context features of each semantic unit are obtained through the bidirectional recurrent neural network in the encoding layer, and the forward context features and backward context features are combined into complete context features; In the multi-head attention layer of the deep neural network, based on the complete context features, the association weight distribution between different semantic units is calculated in parallel by multiple attention heads in the multi-head attention layer. In the fusion layer of the deep neural network, the association weight distributions of multiple attention heads are weighted and fused through the fully connected network in the fusion layer to obtain the comprehensive association weight of each semantic unit. In the selection layer of the deep neural network, the comparison unit in the selection layer compares the comprehensive association weight of each semantic unit with a preset weight threshold to filter out semantic units whose comprehensive association weight is greater than the preset weight threshold as key semantic information.

4. The method according to claim 1, characterized in that, The step of performing differential privacy processing on the intermediate representation data based on data sensitivity level and query frequency, and sharing the processed data in the Industrial Internet, includes: Based on the content attributes of the intermediate representation data, determine the data sensitivity level corresponding to the intermediate representation data; The query frequency is obtained by counting the number of historical queries of the intermediate data within a preset time period. Based on the data sensitivity level and the query frequency, the target noise addition parameter is calculated using a differential privacy algorithm; Random noise is added to the intermediate representation data based on the target noise addition parameters. The intermediate representation data after adding random noise will be shared hierarchically on the industrial internet platform according to the access permission level of the recipient.

5. The method according to claim 4, characterized in that, The step of calculating the target noise addition parameters based on the data sensitivity level and the query frequency using a differential privacy algorithm includes: Based on the data sensitivity level, the corresponding initial privacy parameters are obtained by querying the preset data sensitivity mapping table; Based on the distribution range of the query frequencies, determine the frequency influence factor; The initial privacy parameters and the frequency influence factor are input into the privacy computing engine, and intermediate privacy parameters are generated through the parameter synthesis function in the privacy computing engine. Based on the intermediate privacy parameters, the initial noise addition parameters are calculated using the Laplace distribution function; By combining historical performance data, the initial noise addition parameters are optimized and adjusted to obtain the target noise addition parameters.

6. The method according to claim 1, characterized in that, The process of parsing the device data to obtain protocol type characteristics and structured data includes: Identify protocol header information from the device data, determine protocol type characteristics based on the protocol header information, and match the corresponding protocol specification template according to the protocol type characteristics; Based on the protocol specification template, valid fields are extracted from the device data, and the valid fields are organized into structured data in a unified format.

7. The method according to claim 1, characterized in that, The step of dynamically distributing the structured data to edge computing nodes or cloud computing nodes based on the protocol type characteristics, the data volume, and the real-time requirements of the structured data through a load balancing strategy includes: The first path weight is determined based on the protocol type characteristics, the second path weight is determined based on the data volume, and the third path weight is determined based on the real-time level. The first path weight, the second path weight, and the third path weight are input into the path calculation model in the load balancing strategy, and the priority score of the edge computing node and the priority score of the cloud computing node are calculated through the path calculation model. The priority scores of the edge computing nodes and the priority scores of the cloud computing nodes are compared, and the nodes with higher priority scores are identified as candidate target nodes. Obtain the current processing load of the candidate target node; when the current processing load is less than a preset load threshold, distribute the structured data to the candidate target node. When the current processing load of the candidate target node is greater than the preset load threshold, the node with the second highest priority score is selected as the actual target node, and the structured data is distributed to the actual target node. After distribution is complete, update the load status of the edge computing node and the cloud computing node.

8. A data sharing system for the industrial internet, characterized in that, include: The acquisition module is used to acquire device data from various heterogeneous industrial devices in the Industrial Internet. The parsing module is used to parse and process the device data to obtain protocol type characteristics and structured data; The distribution module is used to dynamically distribute the structured data to edge computing nodes or cloud computing nodes based on the characteristics of the protocol type, the data volume and real-time level of the structured data, and through a load balancing strategy. The conversion module is used to perform semantic parsing on the structured data using a deep neural network based on an attention mechanism after the distribution is completed, to obtain key semantic information, and to convert the key semantic information into intermediate representation data. The processing module is used to perform differential privacy processing on the intermediate representation data according to the data sensitivity level and query frequency, and to share the processed data in the Industrial Internet.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the data sharing method for the Industrial Internet as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the data sharing method of the Industrial Internet as described in any one of claims 1 to 7.

Images