A power plant data security processing method

Abstract

The application provides a power plant data security processing method, and belongs to the technical field of power Internet of Things security and privacy computing, which can at least partially solve the problems of complex certificate management of power plant monitoring system, large data aggregation calculation overhead and privacy leakage risk of cloud analysis in the prior art. The application comprises: generating a double-factor private key based on a certificateless public key cryptography system; serializing and encoding multi-dimensional operation data into a single integer using the Chinese remainder theorem; performing homomorphic encryption and signature on the encoded data using the double-factor private key, verifying the signature in batches on the edge side and aggregating in the ciphertext state; loading a neural network model adapted to homomorphic encryption on the cloud to perform ciphertext domain reasoning on the aggregated ciphertext; and decrypting the operation state analysis result on the authorized side. The application realizes ciphertext availability of power plant core data while ensuring data transmission integrity and confidentiality, and reduces communication and calculation overhead.

Description

Technical Field

[0001] This invention belongs to the field of power Internet of Things security and privacy computing technology, specifically relating to a method for secure data processing in power plants. Background Technology

[0002] Currently, power plants are accelerating their transformation towards digitalization and networking. Modern power plants deploy a large number of sensor nodes to monitor operating parameters such as temperature, pressure, vibration, and grid frequency in real time. The power plant environment is characterized by wide node distribution, strong electromagnetic interference, and high data sensitivity, posing a severe challenge to data security.

[0003] Existing data security solutions for power plants often employ traditional public key infrastructure (PKI) systems or bilinear pairing techniques to achieve data aggregation and secure transmission. However, PKI systems rely on complex certificate management mechanisms, incurring significant storage and computational overhead, making them unsuitable for the resource-constrained massive number of sensor nodes within power plants. While bilinear pairing-based computation schemes offer high security, their computational complexity is too high to meet the stringent real-time requirements of high-frequency data acquisition in power plants.

[0004] Furthermore, most existing technical solutions focus on the security of data transmission links, often neglecting the risk of privacy leaks when data is uploaded to the cloud and analyzed by artificial intelligence models. Although homomorphic encryption can support cloud inference in encrypted form and effectively protect data privacy, relying solely on homomorphic encryption is insufficient to simultaneously solve the problem of lightweight access authentication for massive edge nodes, and it is prone to creating transmission and computing bottlenecks when facing large-scale concurrent data.

[0005] Therefore, designing a comprehensive data security processing solution for power plant scenarios that can achieve lightweight access authentication for edge nodes and ensure secure reasoning and analysis of encrypted data in the cloud has become an urgent technical problem to be solved. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art and to provide a method for secure data processing in power plants.

[0007] To achieve the above objectives, the present invention provides a method for secure data processing in power plants, comprising: Based on a certificateless public key cryptography system, the key generation center generates the system master key and public parameters, generates partial private keys for sensor nodes, and the sensor nodes generate auxiliary secret values. The partial private keys and the auxiliary secret values ​​constitute a two-factor private key. Furthermore, multi-dimensional operating data of power plant units are collected, and the Chinese Remainder Theorem is used to map the multi-dimensional operating data into a single integer plaintext. The single integer plaintext is homomorphically encrypted and digitally signed using the two-factor private key to obtain ciphertext and signature; the signature is verified in batches by the edge gateway at the edge side, and the homomorphic property is used to perform algebraic aggregation on the ciphertext of multiple nodes in the ciphertext state to generate aggregated ciphertext; The aggregated ciphertext is input into a pre-trained adaptive homomorphic encryption neural network model, which performs ciphertext-domain forward propagation inference and outputs the encrypted running state prediction result; and The encrypted running state prediction result is decrypted to obtain plaintext running state analysis data.

[0008] Furthermore, the method for generating the partial private key is as follows: The key generation center is selected from the base point on the elliptic curve. Generate system master key and system master public key ,in Let be an integer randomly selected from the integer field of elliptic curve order; The sensor node Generate random numbers As the auxiliary secret value, the computation node's public key is used. and submit the node identifier to the key generation center. With the node's public key ,in This refers to the sequence number of the sensor node; The key generation center calculates the partial private key. ,in For collision-resistant hash functions, The input is the node identifier. With the node's public key splicing; The sensor node Holding the two-factor private key .

[0009] Furthermore, the digital signature and the batch verification are performed in the following manner: During the signature phase, the sensor node Select random number Calculate the commitment value ,in Identify the base point of the elliptic curve; calculate the signature value: ; in For the sensor node The signature value, The random number, For collision-resistant hash functions, For the sensor node Node identifier, For the sensor node The ciphertext, This is the current timestamp. The auxiliary secret value, This refers to the private key; In the batch verification phase, the edge gateway selects a random coefficient vector. ,in To be assigned to the The random coefficients of each sensor node, The total number of sensor nodes participating in the aggregation; verify the equation: ; in For the sensor node The node's public key, The system's master public key, The hash value calculated by the key generation center when generating the partial private key; If the equation is true, the aggregated data is deemed valid.

[0010] Furthermore, the method of mapping the multidimensional operational data into a single integer plaintext using the Chinese Remainder Theorem is as follows: against Dimensional running data vector Select A pairwise coprime prime number , making ,in The number of dimensions of the multidimensional operational data. For the first The numerical values ​​of the dimensional running data, For the first The prime modulus corresponding to the dimension, ; Calculate the total modulus ; Find the unique integer Satisfying the system of congruence equations , ; The integer This serves as the plaintext input for the homomorphic encryption.

[0011] Furthermore, after decrypting the encrypted operational state prediction result, the process also includes the step of restoring the data in each dimension: The decryption integer is obtained by decrypting the execution state prediction result of the encryption using the decryption private key corresponding to the homomorphic encryption. ; For the decrypted integer Calculate in sequence , To restore the first of the multidimensional operational data Independent numerical values ​​of dimension, where For the first The prime modulus corresponding to the dimension, The dimension of the multidimensional running data.

[0012] Furthermore, the algebraic aggregation method is as follows: The edge gateway receiving area The encrypted data uploaded by the sensor nodes ,in This represents the total number of sensor nodes participating in the aggregation. For the first The encrypted data of each sensor node; Calculating aggregated ciphertext using the homomorphic addition property ,in The aggregated ciphertext; The aggregated ciphertext Uploaded to the cloud.

[0013] Furthermore, the neural network model adapted for homomorphic encryption is a neural network built on the Kolmogorov-Arnold network architecture, in which the activation function on each edge is replaced with a linear combination of Legendre polynomials: ; in The replacement activation function, For the input variables of the activation function, For the first The learnable coefficients of order, for Legendre polynomial, Let be the order of the polynomial expansion. index of polynomial order and ; The weight parameters of the neural network are converted into homomorphic encryption format and stored in the cloud. During the inference process, only homomorphic multiplication between the ciphertext and the encrypted weights, as well as homomorphic addition between the ciphertexts, are performed.

[0014] Furthermore, the parallel acceleration method for the ciphertext field forward propagation inference is as follows: The CUDA architecture is used to decompose the large integer operations of the homomorphic encryption into multiple thread blocks; For the basis function expansion operation of the Legendre polynomial, a dedicated kernel function is designed to implement parallel convolution of the polynomial coefficient matrix; Pipeline technology is used to overlap data transmission and computation tasks, reducing the end-to-end latency of forward propagation inference in the ciphertext field.

[0015] Furthermore, the method also includes a blinding factor dynamic update step: Before each data upload cycle, adjacent sensor nodes generate a shared session key using a lightweight key exchange protocol; The shared session key is used to negotiate and update the local random blind factor, so that the algebraic sum of the random blind factors among the adjacent sensor nodes remains constant or changes according to a preset rule. This ensures the correctness of cloud aggregation and decryption while preventing long-term eavesdropping and deduction of single-point historical data.

[0016] Furthermore, the operational status analysis data is used to guide production scheduling or fault early warning. When the predicted indicators in the operational status analysis data exceed a preset threshold, an alarm is triggered and a control adjustment command is generated.

[0017] The beneficial effects of this invention are as follows: This invention employs a certificateless aggregate signature mechanism, eliminating the certificate management overhead in traditional public key infrastructure systems. It also avoids the risk of key escrow based on identity encryption through a two-factor private key structure. Its computational complexity is lower than that of traditional bilinear pairing schemes, making it suitable for sensor nodes in power plants with limited resources.

[0018] This invention combines homomorphic encryption with the Kolmogorov-Arnold network to enable cloud-based reasoning and analysis of power plant operating status without the ability to access plaintext data, thus protecting the privacy of core industrial data and achieving "usable but invisible" data.

[0019] This invention utilizes a GPU parallel computing architecture to accelerate the inference operation of the Kolmogorov-Arnold network in a homomorphic encryption environment, effectively reducing the computational latency of ciphertext domain inference and meeting the needs of real-time monitoring of power plants.

[0020] This invention achieves unified encryption and independent parsing of multidimensional heterogeneous data through serialization encoding technology based on the Chinese Remainder Theorem, thus meeting the monitoring needs of power plants under complex operating conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart of a power plant data security processing method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the system architecture according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the certificateless aggregated signature and ciphertext aggregation process according to an embodiment of the present invention; Figure 4This is a schematic diagram of the KAN neural network inference in the ciphertext field according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0023] The power plant data security processing method of this invention is implemented based on an edge-cloud collaborative architecture. (See also...) Figure 2 The system architecture comprises three layers: a perception layer, a network layer, and an application layer. The perception layer consists of multiple sensor nodes deployed on the generator units, responsible for collecting multi-dimensional physical quantities such as temperature, pressure, and vibration frequency. The network layer consists of edge gateways deployed in various areas of the power plant, responsible for initial data aggregation, authentication, and protocol conversion. The application layer is a cloud-based secure inference platform, deploying a homomorphic inference model based on the Kolmogorov-Arnold network for fault prediction and energy efficiency analysis.

[0024] The sensor nodes in the perception layer employ low-power microcontrollers with built-in elliptic curve cryptography coprocessors, supporting 256-bit elliptic curve scalar multiplication. Each sensor node has local data caching capabilities with a storage capacity of 64KB. The edge gateway in the network layer uses an embedded industrial control computing platform, configured with an ARM Cortex-A72 processor, 4GB of memory, and running a Linux operating system, supporting simultaneous processing of data from 200 sensor nodes. The cloud control center in the application layer is configured with an NVIDIA A100 GPU, a 128-core CPU, and 512GB of memory, running a homomorphic encryption library and the KAN neural network inference engine.

[0025] Example 1 This embodiment uses a boiler monitoring scenario in a thermal power plant as its application background. Fifty sensor nodes are deployed in the boiler area to collect operational data in two dimensions: water-cooled wall temperature and steam pressure. (See reference...) Figure 1 The data security processing method in this embodiment includes the following steps.

[0026] Step S1: System initialization and key distribution.

[0027] Select elliptic curve As a base group, among which It is a 256-bit prime number. A base point is selected on the elliptic curve. Its order is a large prime number. The key generation center randomly selects an integer. As the system master key, calculate the system master public key. The key generation center publishes the system's public parameter set. ,in and These are two independent collision-resistant hash functions. , .

[0028] 50 sensor nodes in the boiler area Perform the following operations respectively. Sensor Node Generate random numbers locally This serves as a self-generated private key component, i.e., an auxiliary secret value. Sensor node. Compute node public key Sensor nodes Node identifier and node public key The data is sent to the key generation center via a secure channel. The key generation center calculates the hash value. Generate a partial private key And return to the sensor node via a secure channel. Sensor nodes Verify the correctness of a portion of the private key and check the equation. Is it valid? If the verification passes, the sensor node... Use the two-factor private key Securely stored locally.

[0029] In this embodiment, all 50 sensor nodes completed the registration process described above. Each node holds an independent two-factor private key, which does not rely on traditional certificates and does not involve key escrow issues. The key generation center only participates in the distribution of a portion of the private keys during the initialization phase; its online participation is not required during subsequent data transmission.

[0030] Step S2: Multidimensional data acquisition and serialization encoding.

[0031] sensor nodes The boiler water-cooled wall temperature value was collected at the current sampling time. and steam pressure value In this embodiment, That is, the running data is a two-dimensional vector. .

[0032] Choose two coprime prime numbers and Because the normal operating temperature range of the water-cooled wall is... ℃, the normal operating range of steam pressure is MPa, the selected prime numbers satisfy and Requirements.

[0033] Calculate the total modulus .

[0034] With sensor nodes For example, the temperature values ​​it collected. ℃, pressure value MPa. Solving for unique integers using the Chinese Remainder Theorem. , so that: ; ; The solution process is as follows. First, calculate... , Then calculate. mold Multiplicative inverse ,as well as mold Multiplicative inverse Final calculation The integer That is, sensor node Plaintext data to be encrypted.

[0035] The remaining 49 sensor nodes, in the same manner, serialized and encoded their respective collected two-dimensional operational data into corresponding single integer plaintext. .

[0036] Step S3: Certificateless aggregated transmission.

[0037] See Figure 3 Each sensor node performs homomorphic encryption and digital signature on the serialized plaintext, and then the edge gateway performs batch verification and ciphertext aggregation.

[0038] sensor nodes Using an additive homomorphic encryption scheme to encrypt plaintext Encryption, obtaining ciphertext Meanwhile, sensor nodes Select random number Calculate the commitment value Sensor nodes Get current timestamp Calculate hash value Calculate the signature value: ; in For sensor nodes The signature value, For the selected random number, Hash value In the model Multiplicative inverse in the sense of multiplication To assist in the secret value, This is a partial private key. Sensor node. data packet Upload to the edge gateway.

[0039] After receiving data packets from 50 sensor nodes within the area, the edge gateway performs batch signature verification. The edge gateway randomly selects a coefficient vector. ,in For small random integers. Edge gateway verification equation: ; Whether it is valid, among which For sensor nodes The node's public key, For the system master public key, This is the hash value calculated by the key generation center when generating a portion of the private key. If the equation holds true, then the aggregated data from all 50 nodes is considered valid.

[0040] By introducing a random coefficient vector, batch verification combines 50 independent signature verification operations into a single aggregate verification operation, reducing the computational complexity of verification from... The number of independent elliptic curve scalar multiplication operations is reduced to a constant number of operations, significantly alleviating the computational burden on edge gateways.

[0041] After successful verification, the edge gateway utilizes the homomorphic addition property to directly perform algebraic aggregation on the ciphertext of 50 nodes in the ciphertext state: ; The aggregated ciphertext It has the following properties: The decrypted value is equal to the sum of all plaintext. The edge gateway will aggregate the ciphertext. Uploaded to the cloud control center.

[0042] Step S4: Secure reasoning in the ciphertext field.

[0043] See Figure 4 The cloud control center loads a pre-trained homomorphic encrypted neural network model based on the Kolmogorov-Arnold network architecture. This model is designed for the operation scenario of boilers in thermal power plants, with aggregated ciphertext as input and encrypted operation status prediction results as output.

[0044] The Kolmogorov-Arnold network differs from traditional multilayer perceptrons in that it places the learnable activation function on the edges of the network, rather than on the nodes. In this embodiment, the KAN network comprises an input layer, two hidden layers, and an output layer. The input layer has two nodes, corresponding to the two dimensions of temperature and pressure. The first hidden layer contains eight nodes, and the second hidden layer contains four nodes. The output layer contains one node, outputting the coking risk index.

[0045] To adapt to the computational characteristics of homomorphic encryption environments, the activation function on each edge of the KAN network is replaced with a linear combination of Legendre polynomials: ; in The replacement activation function, For the input variables of the activation function, For the first The learnable coefficients of order, for Legendre polynomial, The order of the polynomial expansion. In this embodiment, the order of the polynomial expansion is... The value is set to 5. The recurrence relation of the Legendre polynomial is: ; in For zero-order Legendre polynomials, For a first-order Legendre polynomial, For the first Legendre polynomial, and The first Rank and first Legendre polynomials. Since Legendre polynomials only contain addition and multiplication operations, they are naturally compatible with the types of operations supported by homomorphic encryption, avoiding the deep circuit requirements brought about by traditional nonlinear activation functions.

[0046] The model training phase is completed in a plaintext environment. After training, all weight parameters are... The data is converted to a homomorphic encryption format and stored in the cloud. During the inference phase, the cloud only performs homomorphic multiplication of the ciphertext and encryption weights, as well as homomorphic addition between ciphertexts, without performing any decryption operations, thus ensuring dual privacy of data and model parameters.

[0047] The cloud leverages the Tensor Cores of the NVIDIA A100 GPU to accelerate homomorphic operations. Utilizing the CUDA architecture, large integer operations in homomorphic encryption are decomposed into multiple thread blocks for parallel execution. For the basis function expansion operations of Legendre polynomials, a dedicated GPU kernel function is designed to implement parallel convolution operations on the polynomial coefficient matrix. CUDA pipeline technology is employed to overlap data transmission and computation tasks, reducing end-to-end inference latency. In this embodiment, the latency for a single inference iteration is controlled to within 50 milliseconds.

[0048] The cloud will aggregate the encrypted text The input to the KAN network model propagates forward through the ciphertext domains of the input layer, the first hidden layer, the second hidden layer, and the output layer. In each layer, the ciphertext is homomorphically multiplied with the corresponding encryption weights, and the results of each branch are homomorphically added to the next layer node. Finally, the output layer generates the encryption clustering risk index. .

[0049] Step S5: Decryption and Feedback of Results.

[0050] The authorized client is deployed in the power plant's main control room and holds the private key corresponding to the homomorphic encryption scheme. The authorized client receives the encrypted inference results returned from the cloud. Using the private key, a decryption operation is performed to obtain the plaintext coagulation risk index value.

[0051] If the coking risk index exceeds a preset threshold, the system automatically triggers an alarm and generates a control adjustment command to adjust the combustion control strategy. If the coking risk index is within the normal range, the system records the result in the operation log for maintenance personnel to review.

[0052] Through the above steps S1 to S5, this embodiment verifies that the present invention can support efficient cloud AI analysis while ensuring the secure transmission of power plant data, and the computational overhead is significantly lower than that of traditional solutions, effectively solving the problems of cumbersome certificate management and cloud privacy leakage.

[0053] Example 2 This embodiment uses the safety monitoring scenario of a hydroelectric power plant dam as the application background. Eighty sensor nodes are deployed in the dam area to collect operational data in three dimensions: dam displacement, seepage flow, and water level.

[0054] Step S1: System initialization and key distribution.

[0055] Select elliptic curve As a base group, among which It is a 256-bit prime number. A base point is selected on the elliptic curve. Its order is a large prime number. The key generation center randomly selects an integer. As the system master key, calculate the system master public key. The key generation center publishes the system's public parameter set. .

[0056] 80 sensor nodes in the dam area Perform the registration process separately for each sensor node. Generate random numbers locally As a self-generated private key component. Sensor node Compute node public key Sensor nodes Node identifier and node public key The data is sent to the key generation center via a secure channel. The key generation center calculates the hash value. Generate a partial private key And return to the sensor node via a secure channel. Sensor nodes Verify the equation If the verification is successful, the two-factor private key is stored. .

[0057] Step S2: Multidimensional data acquisition and serialization encoding.

[0058] sensor nodes Collect dam displacement values ​​at the current sampling time. seepage flow value and water level value In this embodiment, That is, the running data is a three-dimensional vector. .

[0059] Choose three pairwise coprime prime numbers , and Because the normal monitoring range for dam displacement is... mm, the normal monitoring range for seepage flow is L / min, the normal monitoring range for water level is m is a prime number chosen to satisfy the range of values ​​for each dimension of the data.

[0060] Calculate the total modulus .

[0061] With sensor nodes For example, the dam displacement values ​​it collected. mm, seepage flow rate L / min, water level value m. Solving for unique integers using the Chinese Remainder Theorem. , so that: ; ; ; The solution process is similar to that in Example 1. First, calculate... , , Then calculate each mold Multiplicative inverse Final calculation The integer That is, sensor node Plaintext data to be encrypted.

[0062] The remaining 79 sensor nodes serialized and encoded their respective collected 3D operational data into corresponding single integer plaintext in the same way.

[0063] Step S3: Certificateless aggregated transmission.

[0064] Each sensor node performs homomorphic encryption and digital signature on the serialized plaintext. Sensor Node Using an additive homomorphic encryption scheme to encrypt plaintext Encryption yields ciphertext Sensor nodes Select random number Calculate the commitment value Sensor nodes Get current timestamp Calculate hash value Calculate the signature value ,in Hash value In the model Multiplicative inverse in the true sense. Sensor node. data packet Upload to the edge gateway.

[0065] After receiving data packets from 80 sensor nodes within the area, the edge gateway performs batch signature verification. The edge gateway randomly selects a coefficient vector. Edge gateway verification aggregate signature equation: ; If the condition is met, then all data from the 80 nodes is considered valid.

[0066] After successful verification, the edge gateway utilizes the homomorphic addition property to perform algebraic aggregation on the ciphertext of 80 nodes in the ciphertext state, and calculates the aggregated ciphertext. The edge gateway will aggregate the ciphertext. Uploaded to the cloud control center.

[0067] Step S4: Secure reasoning in the ciphertext field.

[0068] The cloud control center loads a homomorphic encrypted neural network model based on the Kolmogorov-Arnold network architecture, pre-trained for dam safety monitoring scenarios. The model has three input layer nodes, corresponding to dam displacement, seepage flow, and water level. The first hidden layer contains 12 nodes, and the second hidden layer contains 6 nodes. The output layer contains two nodes, outputting the structural safety risk index and the leakage risk index, respectively.

[0069] In the KAN network, the activation function on each edge is replaced with a linear combination of Legendre polynomials. ,in The replacement activation function, For the first The learnable coefficients of order, for Legendre polynomials. Polynomial expansion order. Set to 6. All weight parameters After training, the data is converted to homomorphic encryption format and stored in the cloud.

[0070] The cloud-based system utilizes an NVIDIA A100 GPU to perform ciphertext-domain forward propagation inference. The CUDA architecture is used to decompose large integer operations in homomorphic encryption into multiple thread blocks for parallel execution. A dedicated kernel function is designed to implement parallel convolution for the basis function expansion operation of Legendre polynomials. Pipeline technology is employed to overlap data transmission and computation tasks. In this embodiment, because the network size with three-dimensional input and dual outputs is larger than in Embodiment 1, the single inference latency is controlled within 80 milliseconds.

[0071] Step S5: Decryption and Feedback of Results.

[0072] Authorized clients are deployed at the dam monitoring center and hold the private key corresponding to the homomorphic encryption scheme. The authorized clients receive the encrypted inference results returned from the cloud, use the private key to perform decryption, and obtain the plaintext structural security risk index and leakage risk index.

[0073] Simultaneously, the authorized client performs the inverse operation of the Chinese Remainder Theorem on the decrypted aggregated integer to restore the independent values ​​of each dimension. (This refers to the process of applying the Chinese Remainder Theorem to the decrypted integer.) Calculate in sequence , , ,in , , Using the prime modulus selected in step S2, the actual values ​​of dam displacement, seepage flow, and water level are restored respectively.

[0074] If the structural safety risk index or leakage risk index exceeds the preset threshold, the system will automatically trigger an alarm and generate a control adjustment command. If both indicators are within the normal range, the system will record the results in the operation log.

[0075] Example 3 This embodiment uses a new energy wind power plant scenario as the application background and introduces a blind factor dynamic update mechanism to enhance forward security. The wind farm deploys 100 sensor nodes to collect operational data in four dimensions: blade vibration frequency, generator winding temperature, wind speed, and rotational speed.

[0076] Step S1: System initialization and key distribution.

[0077] Select elliptic curve As a base group, among which It is a 256-bit prime number. A base point is selected on the elliptic curve. Its order is a large prime number. The key generation center randomly selects an integer. As the system master key, calculate the system master public key. The key generation center publishes the system's public parameter set. .

[0078] 100 sensor nodes in a wind farm Perform the registration process separately for each sensor node. Generate random numbers locally As a self-generated private key component, compute the node's public key. and submit the node identifier to the key generation center. With node public key The key generation center calculates the hash value. Generate a partial private key Return to sensor node Sensor nodes Verify the equation If the verification is successful, the two-factor private key is stored. .

[0079] Step S2: Multidimensional data acquisition and serialization encoding.

[0080] sensor nodes Collect the blade vibration frequency value at the current sampling time. Generator winding temperature value Wind speed value and speed value In this embodiment, That is, the data being processed is a four-dimensional vector. .

[0081] Select four pairwise coprime prime numbers , , and Calculate the total modulus. .

[0082] With sensor nodes For example, the collected blade vibration frequency values Hz, generator winding temperature value ℃, wind speed value m / s, rotational speed rpm. Solving for unique integers using the Chinese Remainder Theorem. , so that: ; ; ; ; The solution process is similar to the previous embodiments, and each step is calculated separately. and its model Multiplicative inverse Final calculation The remaining 99 sensor nodes encode their respective collected four-dimensional operational data into a single integer plaintext in the same way.

[0083] Blinding factor dynamic update steps.

[0084] Before each data upload cycle, adjacent sensor nodes perform dynamic updates to the blind factor. Sensor Nodes Its neighboring nodes Generate a shared session key using the Diffie-Hellman key exchange protocol. ,in For sensor nodes With sensor nodes The session key negotiated between the sensor nodes. Using shared session keys New blind factors are generated using pseudo-random functions. ,in It is a pseudo-random function. This is the current upload cycle number. Sensor node. Generate corresponding blind factors using the same shared session key .

[0085] because The algebraic sum of the blind factors between adjacent nodes remains constant. During the encryption process, sensor nodes... Blinding factor Superimposed on the plaintext, the ciphertext transmitted by the same node at different periods exhibits different randomization characteristics. When the ciphertexts from all nodes are aggregated at the edge gateway, the blinding factors cancel each other out, without affecting the correctness of the aggregation result. This mechanism prevents adversaries from deducing historical data from a single point through long-term eavesdropping, thus achieving forward security.

[0086] Step S3: Certificateless aggregated transmission.

[0087] Each sensor node performs homomorphic encryption and digital signature on the plaintext after adding a blinding factor. Using an additive homomorphic encryption scheme to process plaintext after adding blind factors Encryption yields ciphertext Sensor nodes Select random number Calculate the commitment value Calculate hash value Calculate the signature value Sensor nodes Upload the data packet to the edge gateway.

[0088] After receiving data packets from 100 sensor nodes, the edge gateway performs batch signature verification. The edge gateway randomly selects a coefficient vector. Verify the aggregate signature equation: ; If the condition is met, all data from the 100 nodes is considered valid. After successful verification, the aggregated ciphertext is calculated. Since the algebraic sum of the blind factors is zero, the aggregated ciphertext, after decryption, yields... The results are consistent with those obtained without using a blinding factor. The edge gateway uploads the aggregated ciphertext to the cloud control center.

[0089] Step S4: Secure reasoning in the ciphertext field.

[0090] The cloud control center loads a pre-trained KAN homomorphic encrypted neural network model specifically for wind power generation scenarios. The model has four input layer nodes, corresponding to four dimensions: blade vibration frequency, winding temperature, wind speed, and rotational speed. The first hidden layer contains 16 nodes, the second hidden layer contains 8 nodes, and the third hidden layer contains 4 nodes. The output layer contains 3 nodes, outputting the blade fatigue risk index, the generator overheating risk index, and the predicted wind turbine efficiency, respectively.

[0091] In the KAN network, the activation function on each edge is replaced with a linear combination of Legendre polynomials. ,in The replacement activation function, For the first The learnable coefficients of order, for Legendre polynomials. Polynomial expansion order. Set to 7. All weight parameters are converted to homomorphic encrypted format and stored in the cloud after training.

[0092] The cloud-based NVIDIA A100 GPU performs ciphertext domain forward propagation inference. Since the network size with four-dimensional input and three hidden layers is larger than in the previous embodiment, the CUDA architecture is used to decompose homomorphic operations into more thread blocks for parallel execution. A dedicated kernel function is designed for the basis function expansion operation of Legendre polynomials, and pipelined technology is used to overlap data transmission and computation. In this embodiment, the latency of a single inference operation is controlled within 120 milliseconds.

[0093] Step S5: Decryption and Feedback of Results.

[0094] The authorized client is deployed at the wind farm's central control center and holds the private key corresponding to the homomorphic encryption scheme. The authorized client receives the encrypted inference results returned from the cloud, performs decryption using the private key, and obtains the plaintext blade fatigue risk index, generator overheating risk index, and wind turbine efficiency prediction value.

[0095] The authorized client performs the inverse operation of the Chinese Remainder Theorem on the decrypted aggregate integer, and performs the decrypted integer... Calculate in sequence , , , ,in , , , Using the prime modulus selected in step S2, the actual values ​​of blade vibration frequency, winding temperature, wind speed, and rotational speed are restored respectively.

[0096] If the blade fatigue risk index or the generator overheating risk index exceeds the preset threshold, the system will automatically trigger an alarm and generate control adjustment commands to reduce the turbine speed or activate the cooling system. If all indicators are within the normal range, the system will record the results in the operation log.

[0097] In summary, the embodiments of the present invention have at least the following technical effects: This invention eliminates the overhead of traditional certificate management through a certificateless aggregate signature mechanism, and avoids key escrow risks through a two-factor private key structure, making it suitable for sensor nodes in power plants with limited resources.

[0098] This invention combines homomorphic encryption with the Kolmogorov-Arnold network to achieve cloud-based ciphertext domain inference, thus protecting the privacy of core industrial data.

[0099] This invention utilizes GPU parallel computing architecture to accelerate homomorphic operations, effectively reducing the inference latency of the ciphertext field.

[0100] This invention achieves unified encryption and independent parsing of multidimensional heterogeneous data through serialization encoding based on the Chinese Remainder Theorem.

[0101] This invention achieves forward security through a blind factor dynamic update mechanism, preventing adversaries from deducing single-point historical data.

[0102] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for secure data processing in power plants, characterized in that, include: Based on a certificateless public key cryptography system, the key generation center generates the system master key and public parameters, generates a partial private key for the sensor node, and the sensor node generates an auxiliary secret value. The partial private key and the auxiliary secret value constitute a two-factor private key. Furthermore, multi-dimensional operating data of power plant units are collected, and the Chinese Remainder Theorem is used to map the multi-dimensional operating data into a single integer plaintext. Using the two-factor private key, homomorphic encryption and digital signature are performed on the single integer plaintext to obtain ciphertext and signature; The signature is verified in batches at the edge by the edge gateway, and the homomorphic property is used to perform algebraic aggregation on the ciphertext of multiple nodes in the ciphertext state to generate aggregated ciphertext; The aggregated ciphertext is input into a pre-trained neural network model for adaptive homomorphic encryption, which performs ciphertext domain forward propagation inference and outputs the encrypted running state prediction result. as well as The encrypted running state prediction result is decrypted to obtain plaintext running state analysis data.

2. The power plant data security processing method according to claim 1, characterized in that, The method for generating the aforementioned private key is as follows: The key generation center is selected from the base point on the elliptic curve. Generate system master key and system master public key ,in Let be an integer randomly selected from the integer field of elliptic curve order; The sensor node Generate random numbers As the auxiliary secret value, the computation node's public key is used. and submit the node identifier to the key generation center. With the node's public key ,in This refers to the sequence number of the sensor node; The key generation center calculates the partial private key. ,in For collision-resistant hash functions, The input is the node identifier. With the node's public key splicing; The sensor node Holding the two-factor private key .

3. The power plant data security processing method according to claim 2, characterized in that, The digital signature and the batch verification are performed as follows: During the signature phase, the sensor node Select random number Calculate the commitment value ,in It is the base point of the elliptic curve; Calculate the signature value: ； in For the sensor node The signature value, The random number, For collision-resistant hash functions, For the sensor node Node identifier, For the sensor node The ciphertext, This is the current timestamp. The auxiliary secret value, This refers to the private key; In the batch verification phase, the edge gateway selects a random coefficient vector. ,in To be assigned to the The random coefficients of each sensor node, The total number of sensor nodes participating in the aggregation; verify the equation: ； in For the sensor node The node's public key, The system's master public key, The hash value calculated by the key generation center when generating the partial private key; If the equation is true, the aggregated data is deemed valid.

4. The power plant data security processing method according to claim 1, characterized in that, The method of mapping the multidimensional operational data into single integer plaintext using the Chinese Remainder Theorem is as follows: against Dimensional running data vector Select A pairwise coprime prime number , making ,in The number of dimensions of the multidimensional operational data. For the first The numerical values ​​of the dimensional running data, For the first The prime modulus corresponding to the dimension, ; Calculate the total modulus ; Find the unique integer Satisfying the system of congruence equations , ; The integer The plaintext is used as the input for the homomorphic encryption.

5. The power plant data security processing method according to claim 4, characterized in that, After decrypting the encrypted running state prediction results, the process also includes the step of restoring the data in each dimension: The decryption integer is obtained by decrypting the execution state prediction result of the encryption using the decryption private key corresponding to the homomorphic encryption. ; For the decrypted integer Calculate in sequence , To restore the first of the multidimensional operational data Independent numerical values ​​of dimension, where For the first The prime modulus corresponding to the dimension, The dimension of the multidimensional running data.

6. The power plant data security processing method according to claim 1, characterized in that, The algebraic aggregation method is as follows: The edge gateway receiving area The encrypted data uploaded by the sensor nodes ,in This represents the total number of sensor nodes participating in the aggregation. For the first The encrypted data of each sensor node; Calculating aggregated ciphertext using the homomorphic addition property ,in The aggregated ciphertext; The aggregated ciphertext Uploaded to the cloud.

7. The power plant data security processing method according to claim 1, characterized in that, The neural network model adapted for homomorphic encryption is a neural network built on the Kolmogorov-Arnold network architecture, in which the activation function on each edge is replaced with a linear combination of Legendre polynomials: ； in The replacement activation function, For the input variables of the activation function, For the first The learnable coefficients of order, for Legendre polynomials Let be the order of the polynomial expansion. index of polynomial order and ; The weight parameters of the neural network are converted into homomorphic encryption format and stored in the cloud. During the inference process, only homomorphic multiplication between the ciphertext and the encrypted weights, as well as homomorphic addition between the ciphertexts, are performed.

8. The power plant data security processing method according to claim 7, characterized in that, The parallel acceleration method for the ciphertext field forward propagation inference is as follows: The CUDA architecture is used to decompose the large integer operations of the homomorphic encryption into multiple thread blocks; For the basis function expansion operation of the Legendre polynomial, a dedicated kernel function is designed to implement parallel convolution of the polynomial coefficient matrix; Pipeline technology is used to overlap data transmission and computation tasks, reducing the end-to-end latency of forward propagation inference in the ciphertext field.

9. The power plant data security processing method according to claim 1, characterized in that, It also includes the step of dynamically updating the blinding factor: Before each data upload cycle, adjacent sensor nodes generate a shared session key using a lightweight key exchange protocol; The shared session key is used to negotiate and update the local random blind factor, so that the algebraic sum of the random blind factors among the adjacent sensor nodes remains constant or changes according to a preset rule. This ensures the correctness of cloud aggregation and decryption while preventing long-term eavesdropping and deduction of single-point historical data.

10. The power plant data security processing method according to any one of claims 1 to 9, characterized in that, The operational status analysis data is used to guide production scheduling or fault early warning. When the predicted indicators in the operational status analysis data exceed a preset threshold, an alarm is triggered and a control adjustment command is generated.

Images