Intelligent industrial internet service system

Abstract

The invention discloses an intelligent industrial internet service system. The system comprises the following modules: a terahertz sensing acquisition module; an edge encryption transmission module; an atlas federation processing module; a twinborn decision application module; and an intelligent operation and maintenance security module. According to the method, feature extraction is performed on different types of data through different algorithms, and deep analysis is performed by using a fusion model and a decision tree algorithm. Abnormal conditions in the production process can be accurately observed, and a detailed and accurate basis is provided for production decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial Internet, and in particular to an intelligent industrial Internet service system. Background Art

[0002] The intelligent industrial Internet service system takes intelligent collaboration as the main body, provides a lightweight and innovative digital capability platform, and builds the base of the factory digital transformation platform.

[0003] However, in the intelligent industrial Internet service system, the automotive manufacturing industry faces many difficult problems. The data relied on for production decision-making is extremely limited. The traditional mode can only obtain a small amount of structured data and cannot integrate multi-modal information such as equipment operation, text records, and images, resulting in difficult accurate positioning of production anomalies, insufficient decision-making basis, and high defective product rates. In addition, the production task allocation lacks scientific planning, and the matching of the computing power of edge computing devices with production tasks is in a disorderly state. There is a lack of an efficient allocation algorithm for task characteristics, equipment performance, and network conditions, resulting in low task processing efficiency, slow response, seriously delaying the production rhythm, and increasing production costs.

[0004] Accordingly, this application proposes an intelligent industrial Internet service system. Summary of the Invention

[0005] The purpose of the present invention is to solve the deficiencies existing in the prior art and propose an intelligent industrial Internet service system.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An intelligent industrial Internet service system includes the following modules:

[0008] Terahertz sensing and acquisition module: Use terahertz sensing to detect the inside of industrial equipment, automatically network and distribute data storage in combination with its own networked blockchain, ensure security and can flexibly increase or decrease nodes, and efficiently collect data;

[0009] Edge encryption transmission module: The edge computing device is equipped with an AI chip, intelligently screens and preprocesses the original data before transmission, and uses quantum encryption during transmission, supporting multi-link parallel and bandwidth dynamic allocation;

[0010] Knowledge graph federated processing module: Establish a knowledge graph based on industrial big data, and use federated learning to realize joint modeling and analysis of multi-enterprise data while protecting privacy;

[0011] Twin decision-making application module: Provide comprehensive decision-making basis, optimize resource allocation, reduce debugging costs, improve supply chain efficiency and accelerate the convergence of decision-making models. At the same time, through the collaboration of various technologies, it can effectively handle complex production scenarios and improve the accuracy and efficiency of overall decision-making;

[0012] Intelligent operation and maintenance security module: The AI algorithm monitors security in real time, resists attacks, adaptively adjusts strategies, predicts faults through machine learning, and dynamically configures resources as needed.

[0013] Preferably, in the terahertz sensing and acquisition module, the terahertz sensing technology uses terahertz waves in a specific frequency band, which can effectively penetrate the outer shells of common industrial equipment and clearly obtain information on internal structures and material properties.

[0014] Preferably, the quantum encryption technology of the edge encryption transmission module is based on a mature quantum key distribution protocol, which can generate unconditionally secure quantum keys to ensure the absolute security of data during multi-link parallel transmission.

[0015] Preferably, when constructing the knowledge graph in the atlas federated processing module, a triple storage structure of entities, relationships, and entities is adopted, which can clearly display the complex relationships in all links of industrial production.

[0016] Preferably, for the multi-modal data fusion and reasoning function of the twin decision-making application module, a convolutional neural network is used to extract image features, and a recurrent neural network is used to process unstructured text to achieve efficient fusion and reasoning of multi-modal data.

[0017] Preferably, for the adaptive distributed architecture optimization function of the twin decision-making application module, according to the real-time complexity of the task and the requirements for response time, the allocation ratio of computing resources between the cloud and the edge is dynamically adjusted.

[0018] Preferably, for the digital twin-driven virtual commissioning function of the twin decision-making application module, the constructed digital twin model has the ability to synchronize the state of physical entities in real time to ensure the accuracy of virtual commissioning results.

[0019] Preferably, for the blockchain-based supply chain collaborative decision-making function of the twin decision-making application module, smart contracts are used to automatically execute collaborative decisions in the supply chain, such as automatically triggering inventory replenishment, production schedule adjustment and other operations.

[0020] Preferably, for the combined function of reinforcement learning and transfer learning of the twin decision-making application module, a parameter transfer method is adopted during transfer learning, and some parameters of the source task model are directly transferred to the target task model to accelerate the convergence of the target task decision model.

[0021] The present invention has the following beneficial effects:

[0022] 1. Feature extraction is performed on different types of data through different algorithms, and then in-depth analysis is carried out using a fusion model and a decision tree algorithm. It can accurately identify abnormal conditions in the production process, clearly distinguish whether it is a deviation in equipment parameters or a quality defect of parts, and provide detailed and accurate basis for production decisions.

[0023] 2. By determining key parameters such as the edge computing device set, task set, and network delay matrix according to the characteristics of production tasks and the status of edge computing devices, the task processing efficiency and response speed are greatly improved, the quality of each production link is improved, thereby shortening the overall production cycle and increasing the production efficiency of the enterprise.

[0024] 3. By constructing a digital twin model covering the entire production process, the production conditions are realistically simulated. In the virtual environment, potential problems such as component interference and unreasonable assembly processes can be accurately discovered in advance. Delays and cost increases in actual production are avoided.

[0025] 4. By building a supply chain collaboration platform based on blockchain technology, enterprises in each link of the supply chain share data on the platform. With the help of smart contracts and linear programming algorithms. When facing new order demands, the smart contract flexibly adjusts the allocation strategy based on real-time data, effectively avoiding the risks of inventory backlog and production delays.

[0026] 5. By constructing a reward function with the orientation of maximizing output and minimizing energy consumption through reinforcement learning, the agent obtains feedback by continuously trying different production parameters, and then optimizes its own strategy. It can quickly find the best combination of production parameters in a new scenario and improve production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a partial code display diagram of multimodal data fusion reasoning in Embodiment 1 of an intelligent industrial Internet service system proposed by the present invention;

[0028] Figure 2 It is a partial code display diagram of adaptive distributed architecture optimization in Embodiment 2 of an intelligent industrial Internet service system proposed by the present invention;

[0029] Figure 3 It is a partial code display diagram of digital twin-driven virtual commissioning in Embodiment 3 of an intelligent industrial Internet service system proposed by the present invention;

[0030] Figure 4 It is a partial code display diagram of blockchain-based supply chain collaborative decision-making in Embodiment 4 of an intelligent industrial Internet service system proposed by the present invention;

[0031] Figure 5 It is a partial code display diagram of the combination of reinforcement learning and transfer learning in Embodiment 5 of an intelligent industrial Internet service system proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0032] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0033] Embodiment 1:

[0034] Terahertz sensing and acquisition module

[0035] The terahertz emitter emits short-pulse terahertz waves to the components. After these waves propagate inside the components and interact with different structures, reflected waves or transmitted waves carrying the internal structure information of the components are generated. These echo waves are received by the terahertz detector, and using the wave equation form of Maxwell's equations: ( is the electric field strength, c is the speed of light), through complex mathematical calculations and signal processing, it is analyzed whether there are defects such as micro-cracks, porosity, and voids inside the components, and the detection accuracy can reach the sub-millimeter level.

[0036] Edge encryption transmission module

[0037] The edge encryption transmission module uses the AES encryption algorithm to ensure the security of data transmission. Taking a 128-bit key as an example, the encryption process is divided into 10 rounds. Each round contains four main steps: byte substitution, row shift, column mixing, and round key addition. Byte substitution replaces each byte with a corresponding new byte by looking up the S-box; row shift cyclically shifts the bytes in each row according to certain rules; column mixing mixes the bytes in each column through matrix operations; round key addition performs an exclusive OR operation on the current round's sub-key and the processed bytes. Through this series of complex operations, the encryption formula C = E(K, P) (C is the ciphertext, E is the encryption algorithm, K is the key, P is the plaintext) is realized, ensuring that the component detection data collected is not stolen or tampered with during the process of being transmitted from the production site to the data center.

[0038] Graph federated processing module

[0039] The graph federated processing module uses graph database technology, such as Neo4j, to integrate the data of the entire automobile manufacturing process. From the CAD drawing data in the automobile design stage, to the component bill of materials (BOM) data and production process data during production, and then to the quality inspection data, etc., are all sorted into a triple form of entity-relationship-entity and stored in the knowledge graph. On the platform, the gradient descent algorithm is used to optimize the shared model. Assuming the model parameters are θ and the loss function is L(θ), the formula for updating the parameters in each round is: (α is the learning rate and t is the number of iterations). In this way, while protecting their own core data privacy, enterprises jointly mine potential strategies to improve automobile production quality. For example, it is found that a certain new material can significantly improve the durability of parts under specific processes.

[0040] Twin decision-making application module (multi-modal data fusion and reasoning)

[0041] On the automobile production line, sensors collect structured data such as the temperature T, pressure P, and rotational speed n of equipment in real time. This data can be represented as vectors. For text data such as equipment maintenance manuals, each word is first converted into a low-dimensional vector through the Word2Vec algorithm to capture the semantic relationships between words, and then input into the TextCNN network. After operations such as convolution and pooling, text feature vectors are extracted. For image data captured by surveillance cameras, it is processed using the classic AlexNet convolutional neural network. This network contains multiple convolutional layers, pooling layers, and fully connected layers. Image features are extracted by sliding convolutional kernels on the images, and finally feature vectors are obtained. Adopt a fusion model (α, β, and γ are weight coefficients), and use the backpropagation algorithm to train the deep learning model by minimizing the mean squared error (MSE) loss function (N is the number of samples, y i is the true value, is the predicted value) to determine the optimal weights. When an abnormal assembly of automobile parts is detected, the fused data is input into the decision tree algorithm. The decision tree calculates the information gain:

[0042]

[0043] where (D is the database, a is the attribute, Ent(D) is the information entropy of the dataset D, and D v is the subset of samples in D with the value v for the attribute a) to select the optimal partitioning attribute, quickly locate the root cause of the problem, such as determining whether it is an abnormal parameter of the assembly equipment or a quality problem of the parts themselves, so as to provide an accurate basis for production decisions, effectively improve product quality, and reduce the defective rate.

[0044] Intelligent operation and maintenance security module

[0045] Use a model that combines LSTM and CNN to monitor the network traffic of the production system in real time. LSTM can effectively process time series data and capture the long-term dependence relationships of the data.

[0046] Input gate: i t = σ(W ii x t + b ii + Whi h t-1 +b hi )

[0047] Forget gate: f t = σ(W if x + b if + W hf h t-1 + b hf )

[0048] Output gate: o t = σ(W io x t + b io + W ho h t-1 + b ho )

[0049] Memory cell: c t = f t ⊙ c t-1 + i t ⊙ tanh(W ic x t + b ic + W hc h t-1 + b hc )

[0050] Hidden state: h t = o t ⊙ tanh(c t )

[0051] Among them, the input gate determines how much of the current input information is retained; the forget gate controls how much of the memory cell from the previous moment is retained; the output gate determines the current output information; the memory cell updates the memory state; and the hidden state is used as the output.

[0052] CNN is used to extract local features of network traffic data. By combining the two, the model can automatically identify network attack behaviors such as DDoS attacks and data theft. At the same time, the SVM algorithm is used to analyze device operation data. For non-linearly separable problems, through the radial basis kernel function K(x i , x j ) = exp(-γ||x i - x j || 2 )(γ is the kernel function parameter) maps the data to a high-dimensional space, so as to accurately predict potential device failures. For example, it can predict that the motor of a certain key production device may overheat in the next week, and arrange maintenance in advance to ensure the stable operation of the production line and reduce the downtime caused by device failures.

[0053] It should be noted that in the comparative examples, in Example 1, the accuracy rate of abnormal assembly positioning reached 95.25%, the undetected rate of quality inspection was only 1.85 pieces, and the generalization error of the decision-making model was as low as 3.45%. In Comparative Example 1, relying only on structured data, the accuracy rate of abnormal assembly positioning was only 70.50%, 8.30 pieces were undetected, and the generalization error was 7.20%. In Comparative Example 2, although structured and image data were combined but not effectively fused for reasoning, the accuracy rate was 80.75%, 5.15 pieces were undetected, and the generalization error was 5.40%. The present invention can fully exploit the value of different data, far exceeding the comparative examples in terms of abnormal positioning, quality inspection, and model stability, providing accurate and efficient decision-making support for automobile production. Specifically, as shown in Table 1:

[0054] Table 1: Performance comparison table of multi-modal data fusion reasoning in automobile production decision-making

[0055]

[0056] Example 2:

[0057] Terahertz sensing acquisition module

[0058] The terahertz sensing acquisition module uses the characteristics that terahertz waves can penetrate non-conductive materials and are sensitive to changes in the microscopic structure of substances to perform non-destructive testing on various automobile parts. The working principle and implementation method are the same as those in Example 1.

[0059] Edge encryption transmission module

[0060] The edge encryption transmission module uses the AES encryption algorithm to ensure data transmission security. Taking a 128-bit key as an example, the encryption process is divided into 10 rounds, which is the same as that in Example 1.

[0061] Graph federated processing module

[0062] The graph federated processing module uses graph database technology, such as Neo4j, to integrate the data of the entire automobile manufacturing process. From the CAD drawing data in the automobile design stage, to the bill of materials (BOM) data of parts, production process data, and then to the quality inspection data during production, etc., are all sorted into a triple form of entity-relationship-entity and stored in the knowledge graph. The working mode is the same as that in Example 1.

[0063] Twin decision-making application module (adaptive distributed architecture optimization)

[0064] During the automobile manufacturing process, production tasks are complex and diverse, with different requirements for computing resources and response times. For example, on the automobile parts production line, there are a large number of image recognition tasks (such as detecting surface defects of parts) and data processing tasks (such as real-time analysis of production parameters).

[0065] Let the set of edge computing devices be E = {e1 , e 2 , …, e m}, These devices are distributed in different production areas, and their computing capabilities are C 1 , C 2 , …, C m . The computing capability index can be the number of floating-point operations per second (FLOPS), etc. The task set is T = {t 1 , t 2 , …, t n}. The computing requirements of the tasks are D 1 , D 2 , …, D n . The computing requirements can be measured by the computing complexity, data volume, etc. of the tasks. The network delay matrix is L = (L ij ) m*n , where L ij represents the network delay of task t j allocated to edge computing device e i . The delay can be obtained in real time through network monitoring tools.

[0066] Solve through the optimization objective function using the Hungarian algorithm to reasonably allocate tasks to edge computing devices. In the automotive painting process, the real-time processing task of the painting quality inspection image has extremely high requirements for response time. Through this algorithm, this task is allocated to an edge computing device close to the painting equipment, which has high computing power and low network delay. In this way, the inspection image can be quickly transmitted to the edge computing device for processing, and the processing result can be timely fed back to the painting equipment, enabling the painting equipment to accurately adjust parameters such as the spray gun pressure and painting flow rate according to real-time data, effectively improving the painting quality, reducing the defective rate, and at the same time improving production efficiency and shortening the production cycle.

[0067] Intelligent operation and maintenance security module

[0068] The intelligent operation and maintenance security module uses a model that combines LSTM and CNN to monitor the network traffic of the production system in real time, identify network attack behaviors, and uses the SVM algorithm to analyze the device operation data to predict potential faults. The working principle and implementation method are the same as those in Embodiment 1.

[0069] It should be noted that in the comparative example, the resource utilization rate of the edge device can reach 85.30%, the real-time task response time is shortened to 49.50 milliseconds, and the average load balance degree of the edge device is as high as 0.92. On the contrary, in Comparative Example 1, all tasks are concentrated in the cloud for processing, resulting in a resource utilization rate of only 30.40% for the edge device, a real-time task response time as long as 200.70 milliseconds, and a load balance degree of 0.32; in Comparative Example 2, tasks are randomly assigned to the edge device, with a resource utilization rate of 50.10%, a response time of 120.40 milliseconds, and a load balance degree of 0.58. Through intelligent task allocation, the present invention is significantly superior to the comparative examples in terms of resource utilization, response time, and load balance, effectively ensuring the efficient and stable operation of automobile production, as shown in Table 2 below:

[0070] Table 2: Comparison table of the impact of the adaptive distributed architecture optimization on the performance of edge devices

[0071]

[0072]

[0073] Example 3:

[0074] Terahertz sensing and acquisition module

[0075] The terahertz sensing and acquisition module uses the characteristics that terahertz waves can penetrate non-conductive materials and are sensitive to changes in the microscopic structure of substances to perform non-destructive testing on various automotive parts. The working principle and implementation method are the same as those in Example 1.

[0076] Edge encryption transmission module

[0077] The edge encryption transmission module uses the AES encryption algorithm to ensure data transmission security. Taking a 128-bit key as an example, the encryption process is divided into 10 rounds, which is the same as that in Example 1.

[0078] Graph federated processing module

[0079] The graph federated processing module uses graph database technology, such as Neo4j, to integrate the data of the entire automobile manufacturing process. From the CAD drawing data in the automobile design stage, to the bill of materials (BOM) data and production process data of parts during production, and then to the quality inspection data, etc., are all sorted into triples in the form of entity-relationship-entity and stored in the knowledge graph. The working mode is the same as that in Example 1.

[0080] Twin decision-making application module (virtual commissioning driven by digital twin)

[0081] Build a digital twin model for the entire automobile production process, covering all links from raw material processing, parts manufacturing to vehicle assembly. By simulating the operating states under different production conditions through simulation, using computational fluid dynamics formulas, such as the continuity equation (where ρ is the fluid density, is the velocity vector), this equation is based on the law of conservation of mass, ensuring that in the production process, such as gas flow during automobile welding, paint mist diffusion during painting, etc., substances will not be created or disappear out of thin air. The Navier-Stokes equation (where p is the pressure, μ is the dynamic viscosity, is the external force), based on the law of conservation of momentum, describes the motion law of the fluid.

[0082] The finite volume method is used to divide the calculation region into multiple control volumes. Integral discretization is performed on each control volume to obtain a discretized equation set, and the key parameter distributions in the automobile production process, such as welding, assembly, etc., such as the welding temperature field distribution, the change of component assembly force, etc., are obtained by iterative solution.

[0083] During the virtual commissioning process, problems such as component interference and unreasonable assembly processes can be discovered in advance. For example, in the R & D stage of a new vehicle model, it is found through virtual commissioning that there is interference between two components during the assembly process, and the design scheme is adjusted in time, avoiding production delays and cost increases caused by design problems in actual production, effectively shortening the production cycle and reducing the R & D cost.

[0084] Intelligent operation and maintenance security module

[0085] The intelligent operation and maintenance security module uses a model that combines LSTM and CNN to monitor the network traffic of the production system in real time, identify network attack behaviors, and uses the SVM algorithm to analyze the device operation data to predict potential faults. The working principle and implementation method are the same as those in Embodiment 1.

[0086] It should be noted that in the comparative example, the number of physical prototype productions can be reduced to 1.90 times, the number of monthly process optimizations reaches 10.10 times, and the production process stability improvement ratio is 40.30%. In Comparative Example 1, traditional physical prototype commissioning is used, and the number of productions is as many as 8.20 times, with only 3.15 optimizations per month, and the stability is improved by 10.20%; in the simple 2D simulation method of Comparative Example 2, the number of productions is 5.05 times, with 5.05 optimizations per month, and the stability is improved by 15.10%. The present invention far exceeds the comparative examples in reducing physical prototype production, improving the frequency and stability of process optimization, greatly improving the R & D and process levels of automobile production, as shown in Table 3 specifically:

[0087] Table 3: Comparison table of supply chain collaborative decision-making efficiency based on blockchain

[0088]

[0089] Example 4

[0090] Terahertz sensing and acquisition module

[0091] The terahertz sensing and acquisition module utilizes the characteristics that terahertz waves can penetrate non-conductive materials and are sensitive to changes in the microscopic structure of substances to perform non-destructive testing on various automotive parts. The working principle and implementation method are the same as those in Embodiment 1.

[0092] Edge encryption transmission module

[0093] The edge encryption transmission module uses the AES encryption algorithm to ensure data transmission security. Taking a 128-bit key as an example, the encryption process is divided into 10 rounds, which is consistent with Embodiment 1.

[0094] Spectral graph federated processing module

[0095] The spectral graph federated processing module uses graph database technology, such as Neo4j, to integrate the data of the entire automotive manufacturing process. From the CAD drawing data in the automotive design stage to the bill of materials (BOM) data of parts and the production process data during production, and then to the quality inspection data, etc., are all sorted into triples in the form of entity-relationship-entity and stored in the knowledge graph. The working mode is the same as that in Embodiment 1.

[0096] Twin decision-making application module (blockchain-based supply chain collaborative decision-making)

[0097] Automotive manufacturing enterprises, their suppliers, and logistics enterprises share data through the blockchain platform. Use smart contracts to achieve automatic supply chain collaboration, such as order allocation smart contracts.

[0098] Let the set of parts suppliers be S = {s 1 , s 2, …, s x}, with inventories I 1 , I 2 ,... I x . The inventory data is updated to the blockchain in real time to ensure the authenticity and immutability of the data. The set of automotive assembly plants is F = {f 1 , f 2 ,…, f y}, with production capacities C 1 , C 2 ,…, C y . The production capacity information is also recorded on the blockchain. The set of orders is O = {o 1 , o 2 ,…, o z}, and the profit of each order is p 1 , p 2 ,…, p z , and the required number of parts is q 1 , q 2 ,…, q z .

[0099] Through an algorithm (c i (where c is the production capacity utilization rate of the assembly plant), considering the component inventory and the production capacity constraints of the assembly plant, use the linear programming algorithm to solve and automatically allocate orders. When an automotive manufacturing enterprise receives a new batch of orders, the smart contract reasonably allocates the orders according to the inventory situation of each supplier and the production capacity of each assembly plant, avoiding inventory backlogs and production delays. If the inventory of a certain component of a certain supplier is insufficient, the smart contract will automatically adjust the order allocation and purchase from other suppliers with inventory to ensure the smooth progress of production, effectively improving the supply chain collaboration efficiency and reducing the supply chain cost.

[0100] Intelligent operation and maintenance security module

[0101] The intelligent operation and maintenance security module uses a model that combines LSTM and CNN to monitor the production system network traffic in real time, identify network attack behaviors, and uses the SVM algorithm to analyze the device operation data to predict potential faults. The working principle and implementation method are the same as those in Embodiment 1.

[0102] Embodiment 5

[0103] Terahertz sensing and acquisition module

[0104] The terahertz sensing and acquisition module uses the characteristics that terahertz waves can penetrate non-conductive materials and are sensitive to changes in the microscopic structure of substances to perform non-destructive testing on various automotive components. The working principle and implementation method are the same as those in Embodiment 1.

[0105] Edge encryption transmission module

[0106] The edge encryption transmission module uses the AES encryption algorithm to ensure data transmission security. Taking a 128-bit key as an example, the encryption process is divided into 10 rounds, which is consistent with Embodiment 1.

[0107] Graph federated processing module

[0108] The graph federated processing module uses graph database technology, such as Neo4j, to integrate the data of the entire automotive manufacturing process. From the CAD drawing data in the automotive design stage, to the component bill of materials (BOM) data, production process data during production, and then to the quality inspection data, etc., are all sorted into a triple form of entity-relationship-entity and stored in the knowledge graph. The working mode is the same as that in Embodiment 1.

[0109] Twin decision-making application module (combination of reinforcement learning and transfer learning)

[0110] During the automotive production process, taking the automotive engine assembly link as an example, use the reinforcement learning algorithm to optimize the production parameters. With the goal of maximizing the automotive production quantity Q and minimizing the energy consumption E, construct a reward function R = αQ - βE (α, β are weight coefficients), and determine the values of α and β according to the actual production requirements and cost considerations.

[0111] By continuously trying different assembly parameters (such as assembly torque, assembly speed, etc.), the agent obtains feedback according to the reward function, adjusts its own strategy, and gradually finds the optimal combination of production parameters.

[0112] When introducing a new automobile production line, transfer learning technology is used to transfer the production strategy π 1 to the new scenario. For example, there are similarities in the engine assembly process between the new production line and the old production line. Based on the successful assembly strategy obtained by reinforcement learning on the old production line (such as the assembly sequence and force control of specific components), through the policy fine-tuning algorithm π 2 = π 1 +Δπ (Δπ is the policy increment adjusted according to the new scenario), quickly adapt to the new task. During the engine assembly process of the new production line, through transfer learning and policy fine-tuning, it is possible to find the optimal assembly parameters suitable for the new production line faster, improve production efficiency, reduce energy consumption, and reduce the scrap rate during the production process.

[0113] Intelligent operation and maintenance security module

[0114] A model combining LSTM and CNN is used to monitor the network traffic of the production system in real time. LSTM can effectively process time series data and capture the long-term dependencies of the data.

[0115] Input gate: i t = σ(W ii x t + b ii + W hi h t-1 + b hi )

[0116] Forget gate: f t = σ(W if x + b if + W hf h t-1 + b hf )

[0117] Output gate: o t = σ(W io x t + b io + W ho h t-1 + b ho )

[0118] Memory cell: c t = f t ⊙ c t-1 + i t ⊙ tanh(W ic xt +b ic +W hc h t-1 +b hc )

[0119] Hidden state: h t =o t ⊙tanh(c t )

[0120] Among them, the input gate determines how much of the current input information is retained; the forget gate controls how much of the memory cell from the previous moment is retained; the output gate determines the current output information; the memory cell updates the memory state; and the hidden state serves as the output.

[0121] The CNN is then used to extract the local features of the network traffic data. By combining the two, the model can automatically identify network attack behaviors such as DDoS attacks and data theft. At the same time, the SVM algorithm is used to analyze the device operation data. For non-linearly separable problems, through the radial basis kernel function K(x i ,x j ) = exp(-γ||x i -x j || 2 )(where γ is the kernel function parameter), the data is mapped to a high-dimensional space, enabling accurate prediction of potential device failures. For example, it can predict that the motor of a certain key production device may overheat within the next week, and maintenance can be arranged in advance to ensure the stable operation of the production line and reduce the production downtime caused by equipment failures.

[0122] Specifically, for example Figure 1 , in the Python environment, with the help of the numpy library, structured, image data, and corresponding labels are generated to simulate multimodal data. The multimodal_fusion function is defined to fuse the data, and the train_test_split is used to divide the dataset. A model is created and trained using RandomForestClassifier, the test set is predicted through the predict method of the model, and the accuracy_score is used to calculate the accuracy to evaluate the performance. Finally, new data fusion prediction is simulated to achieve multimodal data fusion inference and decision-making.

[0123] Specifically, for example Figure 2 , lists and dictionaries are used to simulate edge devices and tasks, and the adaptive_task_allocation function is defined. The task list is traversed, and tasks are allocated to edge devices with sufficient resources and the lowest load, and their states are updated to complete adaptive task allocation. Then, the average resource utilization rate and average load are calculated to evaluate the effect, and device resource recovery is simulated to reflect the adaptive adjustment of the architecture.

[0124] Specifically, for exampleFigure 3 , a dictionary is used to simulate the physical and digital twin models, and the digital twin model initially replicates the physical system parameters. The virtual_debugging function is defined to adjust the digital twin model parameters to simulate process optimization. The parameters before and after optimization are compared, the change difference is calculated, and the simulation is called multiple times through a loop to demonstrate the continuous optimization ability.

[0125] Specifically, such as Figure 4 , the Block class is defined to ensure data integrity and immutability, and the Blockchain class is defined to create the genesis block and provide a method for adding new blocks. An instance is created and data blocks containing supply chain information are added, and the data of the latest block is output to simulate the data recording and sharing of the blockchain in supply chain decision-making.

[0126] Specifically, such as Figure 5 , the gym library is used to create the CartPole-v1 environment, a pre-trained model is created with the PPO algorithm to complete transfer learning, and then a new model is created to load the pre-trained parameters for reinforcement learning training. The evaluate_policy function is used to evaluate the model performance, and the model is saved and loaded for testing to demonstrate the model saving and loading functions.

[0127] As mentioned above, the above are only the preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. An intelligent industrial Internet service system, characterized in that: Includes the following modules: Terahertz sensing and acquisition module: uses terahertz sensing to detect the inside of industrial equipment, combines self-organizing blockchain to automatically form a network and distribute data storage, ensures security, and can flexibly increase or decrease nodes to efficiently collect data; Edge encryption transmission module: The edge computing device is equipped with an AI chip, which intelligently filters and pre-processes the original data before transmission, uses quantum encryption during transmission, and supports multi-link parallelization and dynamic bandwidth allocation; Graph federation processing module: builds knowledge graphs based on industrial big data and uses federated learning to achieve joint modeling and analysis of multi-enterprise data while protecting privacy; Twin decision application module: provides comprehensive decision-making basis, optimizes resource allocation, reduces debugging costs, improves supply chain efficiency and accelerates the convergence of decision models. At the same time, through the collaboration of various technologies, it effectively responds to complex production scenarios and improves the accuracy and efficiency of overall decision-making; Intelligent operation and maintenance security module: AI algorithms monitor security in real time, defend against attacks, adaptively adjust strategies, use machine learning to predict failures, and dynamically configure resources on demand.

2. An intelligent industrial Internet service system according to claim 1, characterized in that: In the terahertz sensing and acquisition module, the terahertz sensing technology uses terahertz waves in a specific frequency band, which can effectively penetrate the outer casing of common industrial equipment and clearly obtain internal structure and material property information.

3. The intelligent industrial Internet service system according to claim 1 is characterized in that: The quantum encryption technology of the edge encryption transmission module is based on a mature quantum key distribution protocol, which can generate unconditionally secure quantum keys to ensure the absolute security of data in multi-link parallel transmission.

4. The intelligent industrial Internet service system according to claim 1 is characterized in that: When constructing the knowledge graph, the graph federation processing module adopts a triple storage structure of entities, relationships, and entities, which can clearly display the complex relationships among various links of industrial production.

5. The intelligent industrial Internet service system according to claim 1 is characterized in that: The multimodal data fusion reasoning function of the twin decision application module adopts convolutional neural network to extract image features and recurrent neural network to process unstructured text, so as to realize efficient fusion reasoning of multimodal data.

6. The intelligent industrial Internet service system according to claim 1 is characterized in that: The adaptive distributed architecture optimization function of the twin decision application module dynamically adjusts the allocation ratio of computing resources in the cloud and the edge according to the real-time complexity of the task and the requirements for response time.

7. The intelligent industrial Internet service system according to claim 1 is characterized in that: The digital twin-driven virtual debugging function of the twin decision application module constructs a digital twin model with the ability to synchronize the state of physical entities in real time, ensuring the accuracy of the virtual debugging results.

8. The intelligent industrial Internet service system according to claim 1 is characterized in that: The blockchain-based supply chain collaborative decision-making function of the twin decision application module uses smart contracts to automatically execute collaborative decisions in the supply chain, such as automatically triggering inventory replenishment, production schedule adjustment and other operations.

9. The intelligent industrial Internet service system according to claim 1, characterized in that: The twin decision application module combines reinforcement learning with transfer learning, and adopts a parameter transfer method during transfer learning to directly transfer some parameters of the source task model to the target task model, thereby accelerating the convergence of the target task decision model.

Images