An industrial data transaction processing method and system supporting privacy protection and a storage medium

Abstract

The application discloses a kind of industrial data transaction processing method, system and storage medium supporting privacy protection, belong to big data processing and privacy computing technical field.The application is aimed at the problems of value evaluation distortion, blind state quality quantization difficulty and side channel attack risk etc existing in complete equipment data transaction with strong physical coupling relationship in industrial internet of things, first, the global ciphertext topology graph is constructed by double-mode physical coupling attenuation model, and the mechanical transmission or space flow field coupling between devices is quantified as a secret state adjacency matrix.Second, the edge node performs normalized entropy value calculation on high-frequency sensing signal, and the blind state quality is identified and innately authorized by smart contract.Then, the spatial centrality and gradient cosine similarity are calculated in graph federation iteration, and the exogenous contribution weight is accumulated.Finally, the data value is comprehensively evaluated by a nonlinear coupling formula, the counterparty share is determined, and the clearing rights and interests are finally issued asynchronously by the confused network with secret state credentials.

Description

Technical Field

[0001] This invention relates to the fields of big data processing and privacy computing technology. Specifically, it relates to an industrial data transaction processing method, system, and storage medium that supports privacy protection in industrial Internet of Things (IIoT) scenarios. It is particularly suitable for the transfer and settlement of data assets of high-value complete sets of industrial equipment with strong physical space and working conditions. Background Technology

[0002] Against the backdrop of the deep integration of the digital economy and the Industrial Internet of Things (IIoT), cross-domain sharing and value transfer of industrial big data have become the core driving force for intelligent manufacturing. Data-driven collaborative modeling for complex industrial equipment, such as remaining life prediction, evolutionary status assessment, and full lifecycle management, has extremely high commercial and engineering value. In real-world IIoT scenarios, large-scale industrial equipment rarely operates in isolation; their individual nodes often influence each other through rigid / flexible transmission of mechanical shafts or interference of spatial dynamic fields, exhibiting strong spatiotemporal and load coupling relationships. Sudden changes or minor anomalies in the operating conditions of upstream node equipment can generate significant cascading effects along the physical topology network, exacerbating alternating loads and fatigue degradation in downstream related equipment. Therefore, to train a high-precision intelligent prediction model, it is essential to effectively integrate massive industrial samples spanning different lifecycles and possessing non-standard variable load characteristics.

[0003] Existing general big data trading frameworks are difficult to apply directly to the industrial sector, primarily due to the unique characteristics of industrial data trading. On one hand, enterprises have extreme confidentiality requirements. High-frequency time-series sensor data from underlying equipment and core control strategies directly expose the actual production capacity saturation, energy efficiency limits, and equipment degradation limits of manufacturing enterprises, leading data owners to refuse any form of plaintext sharing. Simultaneously, the non-standard nature of industrial data makes quantifying the quality of blind-state data extremely difficult, and the value density of massive amounts of normally operating data is extremely low. Traditional privacy trading methods, without decryption, are unable to prevent the inflating of data with inferior data and struggle to accurately identify truly high-value long-tail degradation features. On the other hand, the unique business settlement ecosystem of "data for intelligence" differs from the direct monetization in general data trading. Equipment operators urgently need to leverage their refined high-value anomaly features to exchange for higher-level predictive algorithm licenses or equipment lifespan extension optimization strategies from algorithm developers.

[0004] Currently, privacy-preserving data transaction methods based on standard federated learning and game theory pricing are the closest existing implementations. Existing solutions focus on cryptographic data isolation, and the core transaction and processing steps are as follows:

[0005] S1. Joint Modeling Task Release and Fund Locking. Data requesters release training tasks for complex equipment status prediction models on the trading platform, providing the initial neural network structure, and pre-locking funds for settlement in a smart contract, such as fiat currency or universal digital tokens.

[0006] S2. Isolated Local Training. Each data provider trains the distributed model locally and iteratively within its own secure local environment, utilizing its own historical operational data. In this conventional federated architecture, the various data sources participating in the training are treated as independent "islands," completely ignoring the objective physical topology and coupling relationships between underlying devices within the same local physical scenario (such as a single enterprise factory area). Their data are treated merely as absolutely independent statistical samples, undergoing forward propagation and backward differentiation.

[0007] S3. Upload of Encrypted Parameters and Global Aggregation. Each data provider encrypts or adds differential privacy noise to the model weights or gradients generated during local training before uploading them to the central server. The central server updates the global model by weighted averaging of the gradients from each provider in encrypted form. Subsequently, the central server distributes the updated global model parameters to each data provider. After receiving and synchronizing the global model, each data provider starts a new round of iterative training based on its local data. The above process of "local training - encrypted upload - global aggregation - model distribution" is repeated until the prediction accuracy of the global model reaches a preset threshold or meets the convergence condition.

[0008] S4. Marginal contribution assessment based on Shapley value. After model training is completed, the system evaluates the marginal contribution of each isolated node to the global model by calculating the difference in global model prediction accuracy when each data provider node joins or leaves the federated network in an offline state.

[0009] S5. Real-time peer-to-peer fund settlement. Based on the contribution weights calculated in S4, the smart contract proportionally splits the locked settlement funds in real time and explicitly and one-time transfers them to the accounts of each data provider through the blockchain transfer system.

[0010] However, when existing technologies are applied to big data trading scenarios involving complex industrial IoT equipment, the contradiction between the isolated federated architecture and the reality of strong physical coupling leads to severe distortions in data value assessment. Existing technologies treat the data from each data provider node as statistically independent "islands," calculating marginal contributions solely based on the independent training results and Shapley values ​​of a single node's data. This completely severs the objective spatiotemporal topology and physical coupling relationships (such as coupling between mechanical transmission, flow fields, or dynamic fields) between the underlying equipment within each data provider (i.e., within the same local physical scenario). Consequently, the system cannot identify high-value data located on critical paths of fault contagion or at core nodes of the physical topology, causing data value pricing to deviate from common sense in industrial physics and making it impossible to achieve true quality-based pricing. Summary of the Invention

[0011] 1. Technical problem to be solved:

[0012] In view of the problems existing in the prior art, the purpose of this invention is to provide an industrial data transaction processing method, system and storage medium that supports privacy protection, so as to solve the problems of data value assessment distortion, difficulty in blind state quality quantification and vulnerability to side-channel attacks in the prior art, and realize the secure, fair and efficient transaction of industrial data.

[0013] 2. Technical Solution:

[0014] To solve the above problems, the present invention adopts the following technical solution.

[0015] A privacy-preserving industrial data transaction processing method includes the following steps:

[0016] S1: Joint Modeling Task Release and Ciphertext Topology Graph Construction: The data requester releases a joint modeling task on the aggregation platform and locks settlement assets; each edge data provider node extracts the unique identifier of its local device, generates ciphertext node identifiers using dynamic random salt values ​​and uploads them, constructing a set of vertices for the global topology graph; the edge nodes calculate the physical coupling strength using a dual-modal physical coupling attenuation unified model based on the dominant coupling medium type between devices in the same local physical scene, and generate ciphertext coupling weights and directed edge information using the platform's public key before uploading them; the aggregation platform receives the ciphertext nodes, directed edges, and ciphertext weights, instantiates a global ciphertext adjacency matrix in memory, and fills homomorphic encrypted zero values ​​for cross-domain node pairs without coupling associations;

[0017] S2: Local Feature Extraction and Physical Quality Blind Screening Based on Normalized Entropy: The edge computing gateway collects high-frequency time-series sensor signals from the acquisition device, performs coarse-grained processing using different scale factors, and then performs the extraction based on the embedding dimension. and delay time The sequence is mapped to a high-dimensional phase space; the frequency of the permutation patterns of elements within the phase space vector is statistically analyzed, the probability distribution is calculated, and the normalized entropy value is obtained. And upload it to the blockchain smart contract; the smart contract on the normalized entropy value Blind verification is performed. If the node falls within a preset high-value dynamic range, it is admitted. The comprehensive innate physical quality weight coefficient of the node is calculated by combining the prior weight vector related to the physical mode. ;

[0018] S3: Spatiotemporal Graph Federation Hidden Iteration and Acquired Model Contribution Evaluation: The edge nodes locally calculate local transition probabilities and upload them with homomorphic encryption to construct a dense Markov transition matrix; the aggregation platform calculates the spatial centrality of each node in the global network through dense iterative calculation. The edge nodes that are admitted utilize local data to perform spatiotemporal graph neural network iteration based on the encrypted graph, and encrypt the generated local spatiotemporal graph gradient vectors before uploading them to the platform. The aggregation platform, without decrypting the data, calculates the cosine similarity between the local and global gradients, and combines this with the spatial centrality. Calculate the contribution score for a single round After multiple rounds of iterative model convergence, the accumulated contribution weight coefficients of the acquired model for that node are generated. ;

[0019] S4: Dual-track weighted nonlinear coupling and asynchronous anti-backpropagation liquidation: After model convergence, the blockchain smart contract triggers the liquidation engine, which calculates the comprehensive innate physical mass weight coefficient. Weighting coefficients of contribution from the acquired model Through nonlinear coupling formula Calculate the comprehensive value score And calculate the final transaction consideration share for each node. According to the clearing mode identifier preset by the data requester. ,when At that time, traditional monetization clearing is initiated. At that time, the discrete equity mapping function based on the share threshold is initiated. The process involves clearing up non-monetized tiered intellectual assets; encapsulating the ultimate exclusive rights into encrypted authorization certificates, and distributing them asynchronously in batches through a funds obfuscation network during the end-of-day settlement cycle.

[0020] This invention also provides an industrial data transaction processing system that supports privacy protection, comprising:

[0021] Data demand-side nodes are used to publish joint modeling tasks on the aggregation platform, provide the initial model architecture, and set the settlement mode flag. And lock settlement assets into a blockchain smart contract;

[0022] Edge data provider node clusters are deployed in various industrial production sites, including edge computing gateways, to collect high-frequency time-series sensor signals from equipment, and perform local physical quality identification, encrypted node identifier generation, physical coupling strength calculation and densification, local graph gradient calculation and encrypted uploading.

[0023] The aggregation platform, deployed in the cloud, is used to receive data uploaded by edge nodes, construct a global encrypted topology graph, perform aggregated updates of encrypted parameters, and calculate contribution based on spatiotemporal gradient projection.

[0024] The blockchain and smart contract network is used to record the node admission status, perform blind state verification of physical quality and prior rights based on normalized entropy values, accumulate acquired contributions based on spatial centrality and gradient cosine similarity, trigger a dual-track liquidation engine, and complete asynchronous anti-backward monetization or non-monetization rights liquidation.

[0025] The present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above methods.

[0026] 3. Beneficial effects:

[0027] Compared with the prior art, the technical solution provided by this invention has the following advantages:

[0028] (1) This invention introduces a graph topology sensing mechanism and a bimodal attenuation model. By calculating the physical coupling strength between devices in a local physical scene and encrypting and uploading the data, the system can construct a global encrypted equipment topology map in memory and use a dense weighted algorithm to calculate the spatial centrality of nodes in the global network. This enables the algorithm to perceive the cascading contagion path of device failures or state evolution, allowing devices in physically tightly coupled key positions to obtain a reasonable structural premium, thus correcting the distortion problem of island-style equal distribution.

[0029] (2) This invention pushes the normalized entropy algorithm down to the edge gateway for local execution. By coarsening and phase space reconstruction, it extracts the probability distribution reflecting the nonlinear dynamic evolution trajectory of the underlying device and generates a highly compressed normalized entropy value. The system uses a smart contract to set dynamic threshold ranges for joint blind admission judgment. Stationary invalid data approaches 0, and forged random noise approaches 1. This allows for blind screening of high-value dynamic degradation features without decrypting the original time-series waveform, preventing malicious nodes from using inferior data to fraudulently obtain transaction prices, while also reducing the computational overhead of the global federated graph neural network on invalid data.

[0030] (3) This invention introduces a dual-track clearing engine and a discrete equity mapping function in the clearing process. By identifying the liquidation flag bit Intellectual assets are distributed in a tiered manner based on share thresholds, using a non-linear coupling formula. By decoupling single gradient contributions and encapsulating the final exclusive rights into a cryptographic credential, the data is distributed in batches and asynchronously with obfuscation delayed until the end-of-day settlement cycle. This makes the data flow solution deeply compatible with the technological ecosystem of industrial sites that replaces high-order algorithm models with low-level high-frequency data, while effectively resisting side-channel back-pushing attacks and protecting the company's business privacy.

[0031] It should be noted that the structures not described in this invention are not related to the design points and improvement directions of this invention, and are the same as or can be implemented using existing technologies, so they will not be elaborated here. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention;

[0033] Figure 2 This is an overall execution flowchart of the privacy-protecting industrial data transaction processing method according to an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the encrypted topology mapping and physical quality blind screening of the wind turbine cluster in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of multidimensional weight coupling and tiered equity liquidation in an embodiment of the present invention.

[0036] Terminology Explanation:

[0037] I. Joint Modeling Task Release and Encrypted Topology Graph Construction

[0038] Normalized entropy value;

[0039] , Comprehensive innate physical quality weighting coefficient;

[0040] Spatial centrality;

[0041] Single-round contribution score;

[0042] Discrete equity mapping function;

[0043] For mechanical topological mode indicators;

[0044] It is a spatial field mode indicator factor;

[0045] Structural vibration transmissibility;

[0046] Number of transmission stages;

[0047] Material damping coefficient;

[0048] Reference disturbance output power;

[0049] Euclidean spatial distance;

[0050] Dissipation coefficient;

[0051] The angle between the dominant disturbance direction and the device position vector;

[0052] The probability distribution corresponding to the k-th permutation pattern;

[0053] The normalized entropy calculated for the i-th type of sensor data;

[0054] Single-modal basic weights;

[0055] Local transition probability;

[0056] Dense state transition probability;

[0057] ciphertext node identifier;

[0058] A directed edge from node i to node j;

[0059] The dense-state coupling weight of node i to node j (after encryption);

[0060] Platform public key;

[0061] The physical decay function from node i to node j;

[0062] A. Global ciphertext adjacency matrix;

[0063] The zero value of homomorphic encryption;

[0064] CKKS, a homomorphic encryption scheme that supports approximate floating-point calculations (Cheon-Kim-Kim-Song).

[0065] II. Standardized Entropy Value and Physical Quality Blind Screening Module

[0066] X, High-frequency timing sensing signal sequence;

[0067] L, the length of the high-frequency timing sensing signal sequence;

[0068] s, scale factor;

[0069] The original sensing sequence is coarsened using a scale factor s to generate a multi-scale time series.

[0070] Integer functions;

[0071] Embedding dimension;

[0072] Delay time;

[0073] Mutual information content;

[0074] The kth permutation pattern;

[0075] The frequency of occurrence of the k-th permutation pattern;

[0076] K, the total number of reconstructed vectors at the current scale;

[0077] The permutation entropy value under the scale factor s;

[0078] Natural logarithm function;

[0079] m!, factorial of embedding dimension m;

[0080] I. Number of categories of sensor data;

[0081] The single-modal basic weights of the i-th type of sensor data;

[0082] Prior weight vectors strongly correlated with physical modes;

[0083] III. Contribution Assessment and Dynamic Settlement Module

[0084] t, the iteration round number of graph federated learning;

[0085] , No. Round Local spatiotemporal graph gradient vectors encrypted and uploaded by each edge node;

[0086] , No. The global graph gradient vector updated by round-global aggregation;

[0087] Spatial centrality;

[0088] ,node The set of all downstream adjacent nodes that can have a physical impact;

[0089] Dense Markov transition matrix;

[0090] , fundamental centrality vector;

[0091] Damping coefficient;

[0092] Homomorphic scalar multiplication;

[0093] The contribution weight coefficient of the acquired model;

[0094] The comprehensive value score of the corresponding node i;

[0095] U total Total score;

[0096] V i The liquidation revenue of the i-th edge data provider node;

[0097] Contribution amplification hyperparameter;

[0098] The final transaction consideration share of the i-th node;

[0099] , Preset liquidation mode identifier;

[0100] L1, the first level of intellectual assets;

[0101] L2, the second level of intellectual assets;

[0102] L3, the third level of intellectual assets. Detailed Implementation

[0103] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment uses the joint training of the remaining life prediction model of the main bearing of a wind turbine as an example for illustration, but the application scope of this invention is not limited to this, and it can be applied to data transaction scenarios of various sets of industrial equipment with physical coupling (such as multi-stage rolling mills, compressor units, generator sets, etc.).

[0104] Example

[0105] 1. Example Background

[0106] Suppose the data requester is a large wind turbine manufacturer that needs massive amounts of real bearing vibration data to train a large-scale model for predicting the remaining life of the wind turbine main bearing. The settlement mode is set as follows: This refers to a non-monetary rights-for-data exchange model. Simultaneously, it is assumed that the edge data providers are wind farm D and wind farm F. The basic information of these two wind farms is shown in Table 1.

[0107] Table 1 Basic Information about Wind Farms

[0108]

[0109] 2. System Architecture

[0110] like Figure 1 As shown, the system comprises four core entities:

[0111] Data requester node (wind turbine manufacturer): Responsible for initiating joint modeling tasks, providing the initial model architecture, and setting the settlement mode flag. And lock settlement assets into a blockchain smart contract.

[0112] Edge data provider node cluster (edge ​​computing gateways for wind farms D and F): Deployed at various industrial production sites, directly connected to the underlying complete set of equipment, responsible for collecting high-frequency time-series sensor signals, and performing physical quality identification and local model iteration locally.

[0113] Aggregation platform (central server): responsible for splicing the global ciphertext topology graph, aggregating and updating the ciphertext parameters, and evaluating contribution based on gradients.

[0114] Blockchain and smart contract network: responsible for recording node admission status, executing dynamic pricing logic, and completing asynchronous anti-backward dual-track (monetized and non-monetized) equity settlement.

[0115] 3. Method and Flow (S1~S4)

[0116] This invention provides a privacy-preserving industrial data transaction processing method, the overall execution flow of which is as follows: Figure 2As shown, it includes the following four core steps.

[0117] S1: Joint Modeling Task Release and Ciphertext Topology Graph Construction

[0118] (1) Task Issuance

[0119] Wind turbine manufacturers have released a large-scale model training task for predicting the remaining life of wind turbine main bearings, and have pre-set it in a blockchain-smart contract. The threshold for the rights and interests tier, Level 3 requirements Level 2 requirements Level 1 requirements Meanwhile, the smart contract's equity pool locks L3 level for local life extension control strategy packages, L2 level for high-order fault prediction algorithm API call tokens (limited to 1000 times), and L1 level for basic diagnostic reports.

[0120] (2) Node dense state mapping

[0121] Each edge data provider node extracts the unique identifier (factory serial number) of its local physical device and uses the dynamic random salt value issued by the platform's smart contract. Ciphertext node identifiers are generated using the deterministic secure hash function SHA-256 algorithm:

[0122] ;

[0123] Where v i ID represents the ciphertext node identifier generated after the i-th wind turbine is encrypted and obfuscated. i This represents the manufacturing serial number of the i-th wind turbine, Hash(⋅) represents the hash function used by the system, specifically the SHA-256 hash function, and N represents the total number of physical device nodes in the federated network, which in this embodiment refers to the number of wind turbines. (A total of 5 wind turbines). The aggregation platform receives encrypted node identifiers uploaded by each node and constructs a set of vertices for the global topology graph:

[0124] ;

[0125] Hash truncation and salting operations ensure the irreversible concealment of the device's true identity.

[0126] (3) Construction of dual-modal physical edges

[0127] For the same local physical scenario (wind farm D), aerodynamic field interference exists between equipment. Definition For the node Pointing to node A directed edge. If the physical device Changes in operating conditions can affect equipment in space or through physical media. Then the system determines the edge For cross-domain nodes belonging to different data providers (between wind farms D and F), the system assumes they are physically decoupled and there are no connected edges.

[0128] The physical coupling strength of the edge nodes is calculated using a unified dual-modal physical coupling attenuation model based on the dominant coupling medium type between devices. Since the interaction between WT1 and WT2 / WT3 involves spatial flow field wake interference, a modal indicator factor is set. , The physical attenuation function is:

[0129] ;

[0130] in:

[0131] Characterizes the reference disturbance output power of the device (the disturbance intensity of WT1).

[0132] This is the absolute Euclidean spatial distance between the devices (the distance between WT1 and WT2, WT3).

[0133] It is the dissipation coefficient of a fluid or dynamic field in a specific space;

[0134] This represents the angle between the dominant disturbance direction (wake direction) and the device position vector. Used to forcibly filter out physical interference phenomena in the opposite direction of the flow.

[0135] The edge gateway calculates the following:

[0136] Physical attenuation of WT1 to WT2 ;

[0137] Physical attenuation of WT1 to WT3 ;

[0138] Physical attenuation of WT2 to WT3 ;

[0139] Physical attenuation of WT4 to WT5 ;

[0140] For mechanical shaft system scenarios (such as multi-stage gearboxes), another mode is activated ( , ):

[0141] ;

[0142] in, It represents the structural vibration transmission rate between equipment, which is mainly determined by the coupling stiffness or gear ratio. Indicates the number of transmission stages on the physical topology diagram; The material damping coefficient;

[0143] Edge nodes employ the CKKS homomorphic encryption scheme, which supports floating-point approximation calculations, and utilize the platform's public key. The coupling strength is encrypted to generate dense-state coupling weights. :

[0144] ;

[0145] and ciphertext nodes Directed edge and dense state weights Uploaded to the aggregation platform.

[0146] (4) Global matrix concatenation

[0147] After receiving the data, the graph topology building engine of the aggregation platform instantiates a graph in memory. Ciphertext adjacency matrix :

[0148] ;

[0149] When node and When there is no physical coupling between them, the system fills the corresponding positions in the matrix with homomorphically encrypted zero values:

[0150] ;

[0151] The globally encrypted adjacency matrix generated by the aggregation platform In the middle, the nodes in wind farm D (Corresponding to WT1) extrude directed edges (w) outwards. 12 =0.72 and w 13 =0.65), while the wind turbine nodes of wind farm F and the wind turbine nodes of wind farm D are all filled with homomorphic encrypted zero values, and are in an isolated state (e.g. Figure 3 (As shown in the middle).

[0152] S2: Local Feature Extraction and Blind Screening of Physical Quality Based on Normalized Entropy Values

[0153] (1) Multi-scale coarsening and phase space reconstruction

[0154] Edge computing gateways collect high-frequency time-series sensor signal sequences from underlying equipment. The length is L. This embodiment uses vibration signal as an example, but in practice it can cover various modes such as vibration, three-phase current, temperature, and acoustic emission.

[0155] To extract the degradation features of the device at different physical time scales, edge nodes are first analyzed using a scale factor. The original sequence is coarse-grained to generate multi-scale time series. :

[0156] ;

[0157] in, The value is a preset positive integer, ranging from [1, L], representing different downsampling depths of the data. This indicates rounding down. In this embodiment, it is rounded down. .

[0158] For a given scale factor Edge nodes will coarse-grainedize the sequence. Mapping to a high-dimensional phase space, we reconstruct a set of phase space vectors containing the dynamic evolution trajectory of the physical system:

[0159] ;

[0160] Where m represents the embedding dimension, Indicates the delay time.

[0161] Embedding dimension Physically, it characterizes the state-space degrees of freedom of the underlying nonlinear dynamic system of industrial equipment. Taking into account the computing power constraints of the edge computing gateway deployed in the industrial field and the characteristic complexity of the alternating loads on the mechanical equipment, the system will... Strictly limited to the positive integer range of 3 to 7. This embodiment takes... .like Phase space reconstruction will fail to fully represent the complex alternating loads and degradation states of mechanical equipment, resulting in the loss of high-frequency anomalous features; if Not only the number of permutations and combinations of phase space reconstruction It will cause factorial explosions, leading to overload of computing power and memory overflow at edge nodes, and will also excessively amplify the weak background noise of the system, drowning out the true physical evolution characteristics.

[0162] Delay time The physical observation interval characterizes the phase space expansion of the time-series signal. The algorithm employs an adaptive mutual information method for local dynamic optimization. Specifically, the algorithm evaluates the original sequence... With delayed sequence joint probability distribution Calculate their mutual information The optimal time point is selected as the time point when the mutual information first reaches its minimum value. value:

[0163] ;

[0164] This logic ensures the reconstruction of the vector. The components are neither excessively redundant nor completely independent, thus preserving the continuous evolution characteristics of faults or equipment states. To address the computing power constraints of edge computing nodes, the underlying layer employs a lightweight isometric histogram method for rapid probability density approximation, reducing... The computational complexity is...

[0165] (2) Blind calculation and normalization of permutation entropy

[0166] Edge nodes inside the reconstructed phase space vector The elements are sorted in ascending order of their numerical values. Since the vector length is... The total number of all possible permutations and combinations is Species. Let the first... The permutation pattern is denoted as Edge node statistics show the following at the current scale: Frequency of permutation patterns And calculate the probability distribution of this permutation pattern. Probability The calculation method is frequency. Divide by the total number of reconstructed vectors at the current scale :

[0167] ;

[0168] Among them, the total number of reconstructed vectors The calculation formula is:

[0169] ;

[0170] Calculate the scale factor of edge nodes The permutation entropy value and utilize Normalization is performed to obtain the final normalized entropy value. :

[0171] ;

[0172] In the formula, the normalized entropy value The range of values ​​is .

[0173] In this embodiment, the calculation results for each wind turbine edge gateway are as follows:

[0174] WT1 (Pitting of bearing inner ring, periodic impact): ;

[0175] WT2 (Bearing normal, dynamic change minimal): ;

[0176] WT3 (Loose vibration sensor, Gaussian white noise): ;

[0177] WT4 (early wear of bearing outer ring, weak envelope characteristic): ;

[0178] WT5 (Stable operation, slight load fluctuations): .

[0179] (3) Multimodal joint access and prior weighting

[0180] Considering that edge nodes in industrial settings often simultaneously collect multiple types of heterogeneous sensor data, such as vibration, current, temperature, and acoustic emission data, this invention incorporates a multimodal joint admission and comprehensive weighting mechanism into its smart contract. Assume the edge node submits feature sequences of type I sensor data. For the first... Sensor data ( The normalized entropy of each edge node is calculated. Smart contracts are granted access and authorization according to the following rules:

[0181] (i) Joint admission determination: Smart contract for I Perform a traversal. If at least one exists... Falling into the preset high-value dynamic range, such as If the current multimodal data packet of the node has the potential to reflect the dynamic degradation characteristics of the device, the packet is allowed to enter the station; otherwise, it is rejected.

[0182] In this embodiment, the high-value dynamic access range set by the blockchain-smart contract is [0.30, 0.85] (e.g., Figure 3 (as shown)

[0183] WT1 (0.65) admission, record the corresponding WT1 value. ;

[0184] If WT2 (0.15) is below the lower limit of the range, it will be intercepted and eliminated.

[0185] WT3 (0.98) is higher than the upper limit of the range, so it is judged to be counterfeit or invalid and is intercepted and eliminated;

[0186] WT4 (0.62) admission, record the corresponding WT4 ;

[0187] WT5 (0.35) barely meets the entry threshold; record the corresponding WT5. .

[0188] (ii) Comprehensive Weight Calculation: For multimodal data packets that have been granted access to the site, the smart contract first calculates the weight based on each... Calculate the single-modal fundamental weights Subsequently, a prior weight vector strongly correlated with the physical mode is introduced. ,in Calculate the overall inherent physical quality weighting coefficient of the data packets of this node. :

[0189] ;

[0190] Among them, prior weights The sensitivity of the device state evolution characteristics is preset based on different modal data.

[0191] For example, in the condition monitoring task of an industrial fan gearbox, edge nodes simultaneously acquire three types of modal signals: vibration, stator current, and bearing temperature. Since vibration signals are most sensitive to high-frequency, minute fault impacts, followed by current signals, while temperature signals exhibit significant thermal hysteresis, the system can set a priori weight vector. Suppose that an edge node is in a blind screening process, and the single-mode basic weights corresponding to these three types of signals are respectively... The smart contract calculates the node's comprehensive innate physical quality weight coefficient. .

[0192] For the sake of simplicity, this embodiment uses a single mode (vibration) as an example. Directly equal to each wind turbine value.

[0193] S3: Evaluation of Spatiotemporal Graph Federated Hidden Iteration and Acquired Model Contribution

[0194] (1) Calculation of dense-state weighted centrality

[0195] To replace the traditional Shapley value calculation, which involves exponentially increasing computational costs, this invention employs a dense-state weighted algorithm to calculate spatial centrality. Specifically, it includes the following steps.

[0196] Each edge data provider calculates the physical coupling strength between devices locally. Then, before encrypting and uploading, the node is first calculated in the local plaintext environment. The proportion of influence it exerts on all its downstream nodes, i.e., the probability of local transfer. :

[0197] ;

[0198] in, For nodes The set of all downstream adjacent nodes that can have a physical impact.

[0199] It is important to note that the local transition probability in this system... The calculation relies on the objective physical topology constructed in stage S1, independent of the dynamic data admission results in stage S2. That is, although downstream nodes (such as WT2 and WT3) may be intercepted by the smart contract in stage S2 due to substandard data quality and no longer participate in subsequent global graph model gradient updates, their physical entities and physical interference edges still objectively exist in the physical space. Therefore, these unadmitted nodes will still be retained in the transition probability matrix as topological reference points to ensure that the global spatial centrality of upstream throat nodes (such as WT1) can truly reflect their objective cascading importance in the physical system.

[0200] For example, based on the physical topology preservation mechanism described above, even if WT2 and WT3 are not admitted, the set of physical downstream adjacent nodes of WT1 is still {WT2, WT3}. Therefore, the local transition probability is calculated as follows:

[0201] ;

[0202] Edge nodes use platform public key pairs Perform homomorphic encryption and generate dense state transition probabilities. The data is then uploaded, and the platform constructs a dense Markov transition matrix based on this data. :

[0203] ;

[0204] The aggregation platform initializes the basic centrality vector of all nodes in memory:

[0205] ;

[0206] In a dense environment, the platform uses the cascading contagion logic in the graph topology—where nodes with higher influence are affected by nodes with higher influence and thus have higher influence themselves—to perform iterative calculations using the following equations:

[0207] ;

[0208] in:

[0209] Indicates the first After the next iteration, the node Spatial centrality; The preset damping coefficient, usually taken as 0.85, characterizes the probability that a fault or alternating load will propagate stably along the physical link without random dissipation.

[0210] This represents the homomorphic scalar multiplication operation in the ciphertext state.

[0211] The aggregation platform continues to execute the above iterations until the convergence condition is met:

[0212] ;

[0213] in, For a preset threshold (e.g.) The steady-state value after convergence is the final spatial centrality of that node. .

[0214] In this embodiment, because WT1 is located at a critical position in the wake topology, it achieves a high centrality after convergence. The calculation result is assumed to be... WT4 and WT5 are isolated and have extremely low centrality. Assume they converge to... and .

[0215] (2) Local graph gradient hidden uploading and spatiotemporal gradient cosine projection

[0216] The first time studying in the Federation of Maps In the round of iteration, the trusted batch processing aggregation engine extracts the first... The edge data provider nodes use the CKKS algorithm to encrypt and upload local spatiotemporal graph gradient vectors. And the global graph gradient vector updated after this round of global aggregation. .

[0217] The spatiotemporal graph neural network (ST-GNN) used in this embodiment has a conventional structure, including graph convolutional layers for capturing spatial dependencies and gated recurrent unit (GRU) layers for capturing temporal dependencies. Its specific parameters and training methods can be referred to in the prior art, and will not be repeated here.

[0218] The platform does not decrypt the underlying data; it directly calculates the cosine similarity between the two data points in the parameter space and combines this with the node's position in the encrypted topological graph. Spatial centrality in Calculate the contribution score for a single round. :

[0219] ;

[0220] in, The function is used to filter negative gains, that is, when a node provides a poor or toxic gradient that is opposite to the direction of global optimization, its score for that round is 0.

[0221] In this embodiment, after multiple rounds of federated training, the platform calculates the mean cosine similarity between the local graph gradient and the global gradient. It is assumed that the gradient directions provided by WT1, WT4, and WT5 are all good, with similarities of 0.85, 0.90, and 0.60, respectively.

[0222] go through After the global iterative model converges, the system accumulates the effective contributions over time to generate the final acquired model contribution weight coefficient for that node. :

[0223] ;

[0224] After summation and normalization, we get: .

[0225] S4: Dual-track weighted nonlinear coupling and asynchronous anti-backward liquidation

[0226] (1) Comprehensive value and share assessment

[0227] Once the model converges, the smart contract triggers the liquidation engine to calculate the resulting comprehensive innate physical mass weight coefficients. ( ) and the contribution weights of the above-mentioned acquired model Perform nonlinear coupling and compute nodes Overall value score :

[0228] ;

[0229] in, This embodiment takes the contribution amplification hyperparameter preset by the system. .

[0230] The overall score for each node is calculated as follows:

[0231] Score: ;

[0232] Score: ;

[0233] Score: ;

[0234] Total Score: .

[0235] Subsequently, the node's final share of the transaction consideration in the global federated network is calculated. :

[0236] have to: .

[0237] (2) Dual-track clearing engine execution

[0238] To completely eliminate side-channel back-pushing attacks based on single fund flows and to perfectly meet the diverse business needs of the industry, smart contracts set daily settlement cycles within the system, based on the settlement mode flag set by the data requester when issuing the task. Start the dual-track clearing engine.

[0239] If the liquidation mode flag W=0, the continuous basic settlement mode of "data for currency" is initiated. At this time, the smart contract extracts the total task reward V that the data requester pre-locked in the equity pool when posting the task. total In this model, the smart contract uses a linear proportional division method, directly based on the final consideration share Q of each node. i Calculate the specific revenue, specifically the liquidation revenue V of the i-th edge data provider node. i for:

[0240]

[0241] Assume the total reward locked by the requester for this task is 10,000 Federation tokens. Based on the share calculated by the smart contract for each node, the following linear liquidation will be executed:

[0242] WT1 earnings: One token;

[0243] WT4 earnings: One token;

[0244] WT5 earnings: One token;

[0245] WT2 and WT3 Rewards: Due to failing the quality blind screening, their share is 0, and their reward is 0 tokens. In this embodiment, the data requester sets... This involves adopting a non-monetary equity exchange model of "data for intelligence." In this model, the asset equity pool is transformed into a high-level intellectual asset repository provided by the data requester. Smart contracts no longer perform simple linear partitioning but instead introduce a discrete equity mapping function based on share thresholds. :

[0246] ;

[0247] The system pre-sets four intellectual asset tiers: L1 represents access to basic data analysis reports, L2 represents access to advanced algorithm APIs, and L3 represents local authorization for device life extension optimization strategy packages. Each tier corresponds to a different trigger share threshold. .

[0248] WT1 (59.3%) meets the requirements. This triggers Level 3 and obtains the local life extension control policy package;

[0249] WT4 (31.0%) meets the requirements. Triggering L2 level, obtaining a high-order fault prediction algorithm API call token (limited to 1000 times).

[0250] WT5 (9.7%) meets the requirements. This triggers Level 1 and generates a basic data analysis report.

[0251] WT2 and WT3 were intercepted in the S2 phase, and their final share absolute value is 0, so they will not be liquidated.

[0252] (3) Asynchronous and covert delivery

[0253] After completing the aforementioned settlement, the platform will uniformly encapsulate the calculated final exclusive rights (token shares or tiered intellectual asset authorization codes) into encrypted authorization certificates. During the daily settlement cycle, the system will use a fund obfuscation network based on mixing protocols (such as the CoinJoin mechanism) and zero-knowledge proofs to distribute the funds in batches.

[0254] Specifically, the smart contract packages all payment funds from demanders and receivables from providers into a decentralized obfuscation pool. Within the pool, the direct mapping between input and output addresses is removed, and individual transactions are broken down and mixed through random rearrangement of output order and delay of block confirmation time. This mechanism completely severs the side-channel attack path at the physical execution level, preventing external attackers from using real-time settlement latency, fund flow, or amount to reverse-engineer the gradient contribution or data status of each node.

[0255] This embodiment fully demonstrates the entire process from task deployment, encrypted topology graph construction, normalized entropy blind screening, spatiotemporal graph federation iteration to dual-track step-by-step liquidation. Through the graph topology awareness mechanism, WT1, located at a critical position in the wake, is identified due to its high spatial centrality. Significant structural premiums were achieved; low-quality WT2 and WT3 were intercepted in advance through normalized entropy blind screening; the contributions of each node were transformed into intellectual assets matching their value through dual-track nonlinear coupling and discrete ladder mapping, and the risk of side-channel back-pushing was eliminated through asynchronous obfuscation.

[0256] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art can make various improvements and modifications without departing from the spirit and principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for processing industrial data transactions that supports privacy protection, characterized in that, Includes the following steps: S1: Joint Modeling Task Release and Encrypted Topology Graph Construction: Data requesters release joint modeling tasks on the aggregation platform and lock settlement assets; Each edge data provider node extracts the unique identifier of its local device, generates encrypted node identifiers using dynamic random salt values, and uploads them to construct a set of vertices in the global topology graph; The edge nodes calculate the physical coupling strength using a dual-modal physical coupling attenuation unified model based on the dominant coupling medium type between devices within the same local physical scene, and then upload the dense coupling weights and directed edge information generated using the platform public key. The aggregation platform receives the ciphertext nodes, directed edges, and ciphertext weights, instantiates a global ciphertext adjacency matrix in memory, and fills homomorphic encryption zero values ​​for uncoupled cross-domain node pairs. S2: Local Feature Extraction and Physical Quality Blind Screening Based on Normalized Entropy: The edge computing gateway collects high-frequency time-series sensor signals from the acquisition device, performs coarse-grained processing using different scale factors, and then performs the extraction based on the embedding dimension. and delay time Map the sequence to a high-dimensional phase space; The frequency of the permutation patterns of elements within the phase space vector is statistically analyzed, the probability distribution is calculated, and the normalized entropy value is obtained. And upload it to the blockchain smart contract; the smart contract on the normalized entropy value Blind verification is performed. If the node falls within a preset high-value dynamic range, it is admitted. The comprehensive innate physical quality weight coefficient of the node is calculated by combining the prior weight vector related to the physical mode. ; S3: Spatiotemporal Graph Federation Hidden Iteration and Acquired Model Contribution Evaluation: The edge nodes locally calculate local transition probabilities and upload them with homomorphic encryption to construct a dense Markov transition matrix; the aggregation platform calculates the spatial centrality of each node in the global network through dense iterative calculation. The edge nodes that are admitted use local data to perform spatiotemporal graph neural network iteration based on encrypted graphs, and then encrypt and upload the generated local spatiotemporal graph gradient vectors to the platform. The aggregation platform calculates the cosine similarity between the local and global gradients without decrypting the data, and combines this with the spatial centrality. Calculate the contribution score for a single round After multiple rounds of iterative model convergence, the accumulated contribution weight coefficients of the acquired model for that node are generated. ; S4: Dual-track weighted nonlinear coupling and asynchronous anti-backpropagation liquidation: After model convergence, the blockchain smart contract triggers the liquidation engine, which calculates the comprehensive innate physical mass weight coefficient. Weighting coefficients of contribution from the acquired model Through nonlinear coupling formula Calculate the comprehensive value score And calculate the final transaction consideration share for each node. ; Based on the clearing mode identifier preset by the data requester ,when At that time, traditional monetization clearing is initiated. At that time, the discrete equity mapping function based on the share threshold is initiated. The process involves clearing up non-monetized tiered intellectual assets; encapsulating the ultimate exclusive rights into encrypted authorization certificates, and distributing them asynchronously in batches through a funds obfuscation network during the end-of-day settlement cycle.

2. The method according to claim 1, characterized in that, The unified model for dual-modal physical coupling attenuation in step S1 is expressed as follows: ； in, For mechanical topology modal indicator factor, This is a spatial field modal indicator factor, used when the equipment is connected by mechanical shafts. , When there is interference from spatial flow fields or dynamic fields , ; For structural vibration transmissibility, For the number of transmission stages, The material damping coefficient, As the reference perturbation output power, For Euclidean space distance. Here is the dissipation coefficient. The angle between the dominant disturbance direction and the device position vector.

3. The method according to claim 1, characterized in that, The normalized entropy value in step S2 The calculation method is as follows: the original time series is scaled by a scaling factor. Coarsening is performed to obtain the sequence. Based on embedding dimension and delay time Reconstruct the phase space vectors; Calculate the probability of each permutation pattern occurring. Calculate the permutation entropy and through Normalization; wherein, the embedding dimension The delay time is limited to an integer from 3 to 7. The mutual information method is used to dynamically determine the value. Indicates the first A variety of arrangement patterns.

4. The method according to claim 1, characterized in that, The multimodal joint admission and comprehensive weight calculation in step S2 specifically involves: Let the normalized entropy of the type I sensor data submitted by the edge node be... (where i = 1, 2, 3, ..., I, representing the normalized entropy value calculated for different types of sensor data), if there exists at least one If a data packet falls within the preset high-value dynamic range [0.30, 0.85], then the node's data packet is allowed to enter the station; for multimodal data packets that are allowed to enter the station, the smart contract determines the appropriate data packet based on the specified parameters. Mapping the single-modal fundamental weights And combined with the prior weight vector ,in Calculate the comprehensive innate physical quality weighting coefficient .

5. The method according to claim 1, characterized in that, Spatial centrality in step S3 The calculation is performed using a dense-state weighted algorithm, specifically including: each edge node calculating its local transition probability locally. Homomorphic encryption generates dense state transition probabilities After uploading, the aggregation platform constructs a dense-state Markov transition matrix. Initialize the centrality vector Iterate according to the following formula until convergence: ； in, The damping coefficient is denoted by N, which represents the total number of physical device nodes in the federated network. This represents the homomorphic scalar multiplication operation. For nodes The set of all downstream adjacent nodes that can have a physical impact.

6. The method according to claim 1, characterized in that, The single-round contribution score in step S3 Calculate using the following formula: ； in, For the first Round The local spatiotemporal graph gradient vectors encrypted and uploaded by each edge node. For the first The global graph gradient vector is updated after global aggregation.

7. The method according to claim 1, characterized in that, Discrete interest mapping function in step S4 Specifically: ； L1, L2, and L3 represent local authorization for basic data analysis reports, advanced algorithm API call permissions, and device life extension optimization strategy packages, respectively. , , The preset share threshold and satisfy .

8. The method according to claim 1, characterized in that, The funds obfuscation network in step S4 is implemented using a mechanism based on a coin mixing protocol and zero-knowledge proofs: the smart contract packages all the payment funds of all demanders and the receivables of all providers into a decentralized obfuscation pool, strips the direct mapping relationship between input addresses and output addresses in the pool, and breaks down and mixes single transactions by randomly rearranging the output order and delaying the block confirmation time, thereby realizing the batch asynchronous issuance of encrypted authorization certificates.

9. An industrial data transaction processing system that supports privacy protection, characterized in that, include: Data demand-side nodes are used to publish joint modeling tasks on the aggregation platform, provide the initial model architecture, and set the settlement mode flag. And lock settlement assets into a blockchain smart contract; Edge data provider node clusters are deployed in various industrial production sites, including edge computing gateways, to collect high-frequency time-series sensor signals from equipment, and perform local physical quality identification, encrypted node identifier generation, physical coupling strength calculation and densification, local graph gradient calculation and encrypted uploading. The aggregation platform, deployed in the cloud, is used to receive data uploaded by edge nodes, construct a global encrypted topology graph, perform aggregated updates of encrypted parameters, and calculate contribution based on spatiotemporal gradient projection. The blockchain and smart contract network is used to record the node admission status, perform blind state verification of physical quality and prior rights based on normalized entropy values, accumulate acquired contributions based on spatial centrality and gradient cosine similarity, trigger a dual-track liquidation engine, and complete asynchronous anti-backward monetization or non-monetization rights liquidation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the privacy-protected industrial data transaction processing method according to any one of claims 1 to 8.

Images