A Federated Diagnosis Method and System for Complex Equipment Faults Based on Homomorphic Encryption

By introducing multidimensional physical manifolds and homomorphic gating logic in the ciphertext domain under a fully homomorphic encryption environment, the challenges of data privacy protection and quality control for complex industrial equipment are solved, and the robustness and security of fault diagnosis models for complex equipment are improved.

CN123122784APending Publication Date: 2026-11-17陈志鹏 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611091152.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing encrypted federated learning schemes struggle to balance data privacy protection and data quality control in complex industrial scenarios. In particular, they cannot effectively eliminate anomalous samples that violate physical common sense in fully homomorphic encryption environments, resulting in insufficient robustness and security of fault diagnosis models.

Method used

The design is based on a plaintext self-checking mechanism using multidimensional physical manifolds. It generates physical credentials by utilizing the inherent energy conversion constraints of the equipment and eliminates abnormal samples in the ciphertext state through homomorphic gating logic in the ciphertext field. Combined with fully homomorphic encryption technology, it achieves secure aggregation of the model.

Benefits of technology

Without decrypting the data, it effectively eliminates anomalous samples that violate physical principles, improving the robustness and security of the fault diagnosis model and ensuring a balance between data quality control and privacy protection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123122784A_ABST
    Figure CN123122784A_ABST
Patent Text Reader

Abstract

This invention discloses a federated diagnosis method and system for complex equipment faults based on homomorphic encryption, relating to the field of data security technology. It comprises a trusted third party, an edge monitoring terminal, and a cloud server. The trusted third party is responsible for independent homomorphic key generation and distribution; the edge monitoring terminal, as the data owner, is responsible for collecting data, performing local training and physical consistency verification, and uploading gradients and physical credentials in encrypted form; the cloud server, as the computing core, performs encrypted gating cleaning and model aggregation calculations without obtaining plaintext data and specific physical residual values. Through multidimensional physical manifold construction, encrypted gating generation, and privacy-weighted aggregation, it achieves secure fault diagnosis of complex equipment, solving the problem of balancing data security protection and data quality control in federated learning scenarios, and enabling industrial equipment status monitoring under data security protection.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of data security technology, and more specifically to a method and system for federated diagnosis of complex equipment faults based on homomorphic encryption. Background Technology

[0002] With the rapid development of Industry 4.0 and intelligent manufacturing, monitoring the operational status and diagnosing faults of large and complex equipment (such as generator sets, industrial compressors, and gas turbines) has become a crucial aspect of ensuring production safety. Deep learning technology, due to its powerful feature extraction capabilities, is widely used in equipment fault diagnosis. However, in real-world industrial scenarios, high-quality fault sample data is often scattered across different operating entities or geographical areas (e.g., different wind farms or hydropower stations), forming "data silos." To address this issue, federated learning technology has emerged to jointly train models using multi-party data without sharing the original data.

[0003] Although federated learning protects privacy by exchanging only model parameters (or gradients) rather than the original data, research shows that attackers can still infer the original data or deduce sensitive attributes by analyzing gradient information. Therefore, federated learning combined with fully homomorphic encryption (FHE) has become a research hotspot in the field of privacy computing. It allows cloud servers to perform aggregation operations directly in encrypted form, thereby achieving "data available but not visible".

[0004] However, existing encrypted federated learning schemes face a severe challenge in balancing data privacy protection and data quality control when dealing with complex industrial scenarios. Specifically, complex equipment often operates under harsh conditions, and sensors are prone to drift, failure, or malicious data injection attacks. This "dirty data," generated by equipment malfunctions or attacks, can significantly degrade or even disable the global model's performance if it participates in model training. With the introduction of fully homomorphic encryption, cloud servers cannot decrypt and view the gradients or raw data uploaded by edge nodes, completely depriving them of the ability to monitor and clean data quality. Furthermore, traditional federated learning defense mechanisms rely primarily on the statistical distribution characteristics of gradient data (such as Euclidean distance and cosine similarity) to identify and remove outliers. However, in the ciphertext domain of fully homomorphic encryption, calculating these nonlinear statistical indicators is not only computationally expensive but also makes it difficult to identify deceptive samples whose values ​​are within the statistical range but violate fundamental physical laws, such as data with zero input energy but full output power.

[0005] Therefore, how to strictly protect data privacy and effectively eliminate abnormal samples that violate physical common sense in a fully homomorphic encryption "zero-knowledge" environment, and ensure the robustness and physical consistency of the fault diagnosis model, has become a technical problem that urgently needs to be solved in the field of industrial internet security. Summary of the Invention

[0006] In view of the above problems, this invention proposes a method and system for federated diagnosis of complex equipment faults based on homomorphic encryption to overcome or at least partially solve these problems. It designs a plaintext self-checking mechanism based on a multidimensional physical manifold for the edge computing nodes, generating physical credentials using the inherent energy conversion constraints of the equipment. A novel homomorphic gating logic in the ciphertext domain is designed, mapping the physical credentials to gating weights in the ciphertext state through polynomial approximation technology. This allows the cloud to automatically eliminate anomalous samples that violate physical principles without decrypting the data, significantly improving the robustness and security of the global model in uncontrolled industrial environments. This solves the problem of balancing data security protection and data quality control in federated learning scenarios, enabling industrial equipment status monitoring under data security protection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a federated diagnosis method for complex equipment faults based on homomorphic encryption, comprising the following steps: S1: Each edge monitoring terminal has a pre-installed multi-dimensional physical constraint model for the target equipment. The multi-dimensional physical constraint model includes a baseline physical function relationship and a set physical deviation threshold. The physical deviation threshold is sent to the cloud server. S2: Use a trusted third party to select the CKKS fully homomorphic encryption scheme to generate a key, distribute the public key and private key to the entire network, and store the private key; S3: Each edge monitoring terminal collects real-time operating data of the target equipment, and after preprocessing, constructs a local training sample set. The local training sample set is used to train the local model through federated learning, calculates the gradient update amount of the model weights, and calculates the consistency residual based on the real-time operating data and the relationship between the benchmark physical function. At the same time, the gradient update amount and the physical consistency residual are encrypted and packaged into a ciphertext data packet and sent to the cloud server using the public key. S4: The cloud server uses a key to encrypt the physical deviation threshold to generate an encrypted threshold. Based on the ciphertext data packet and the encrypted threshold, a global ciphertext model is obtained through homomorphic operation and broadcast back to the edge monitoring terminal. S5: The edge monitoring terminal decrypts the global encrypted model to obtain an updated global model. When the global model converges, training stops; otherwise, it returns to S3 to continue iterating through federated learning.

[0009] Preferably, in S1, the edge monitoring terminal establishes an output response by recording the energy conversion characteristics or operating status manifold of the target equipment. Input incentive variables and control state variables The baseline physical function relationship between them .

[0010] Preferably, the multidimensional physical constraint model also includes variable value ranges, manifold constraints of operating states, and dimensional normalization rules.

[0011] Preferably, during the first round of training, each edge monitoring terminal uses the initialization model sent by the cloud server as a local module; the initialization model is a fault diagnosis model built using machine learning or deep learning models, such as ResNet, Transformer or other fault classification / state recognition models.

[0012] Preferably, in S3, each edge monitoring terminal trains its local model using local samples from its local training sample set during each training cycle, and calculates the gradient update amount of the model weights; the updated global model obtained by the edge monitoring terminal during each training cycle is used as the local model.

[0013] Preferably, in S3, each edge monitoring terminal substitutes the physical input parameters and control state variables in the real-time running data into the benchmark physical function relationship in a plaintext environment to obtain the theoretical response, and extracts the actual output response from the real-time running data as the measured response, calculates the absolute difference between the measured response and the theoretical response, and obtains the physical consistency residual.

[0014] Preferably, the actual output response is the actual output data of the target equipment directly collected by the edge monitoring terminal through sensors or obtained after preprocessing, such as actual power, actual efficiency, actual temperature, actual vibration response or other state response quantities; the physical consistency residual is used to determine whether the batch of operating data conforms to the basic physical laws of the target equipment.

[0015] Preferably, in S3, the edge monitoring terminal uses the public key to encrypt the gradient update amount to generate a ciphertext gradient, and encrypts the physical consistency residual to generate a ciphertext physical credential.

[0016] Preferably, the specific process by which the cloud server obtains the global ciphertext model through homomorphic operations based on the ciphertext data packet and the encryption threshold is as follows: S41: The cloud server extracts the ciphertext physical credentials from the ciphertext data packet and uses homomorphic subtraction to calculate the deviation between the ciphertext physical credentials and the encryption threshold of each edge monitoring terminal to obtain the ciphertext deviation. S42: The cloud server normalizes the ciphertext deviation and performs a polynomial approximation algorithm within the ciphertext domain based on the normalized ciphertext deviation to generate ciphertext gating coefficients. If the normalized ciphertext deviation is greater than 0, the ciphertext gating coefficients approach the encrypted 0; if the normalized ciphertext deviation is less than or equal to 0, the ciphertext gating coefficients approach the encrypted 1. S43: The cloud server extracts the ciphertext gradient from the ciphertext data packet and uses homomorphic multiplication to multiply the ciphertext gating coefficients with the ciphertext gradient element by element to obtain the filtered ciphertext gradient. S44: The cloud server uses homomorphic addition to accumulate the gradient of the filtered ciphertext to obtain the global ciphertext aggregation. S45: The cloud server calculates the sum of all ciphertext gating coefficients as the number of effective nodes, and uses homomorphic division or interaction protocol to normalize the global ciphertext aggregation amount to obtain the average gradient update amount. S45: The cloud server updates the current global model based on the average gradient update amount to obtain the updated global encrypted model.

[0017] Preferably, the specific process of S42 is as follows: S421: Use a third-order Chebyshev polynomial to perform coarse screening calculations on the normalized ciphertext bias to obtain the coarse screening gate coefficients. ; S422: Determine the coarse screening gate coefficient If the value is within the valid threshold range, then proceed to fine screening and enter S423. If the coarse screening gate coefficient If the minimum boundary value is less than the effective threshold, then the ciphertext gating coefficient is set to 0; If the coarse screening gate coefficient If the maximum boundary value is greater than the effective threshold, then the coarse screening gate coefficient will be adjusted. Assign the value to the ciphertext gating coefficient; S423: Calculating the screening gating coefficients using a seventh-order Chebyshev polynomial. And call the second-order correction polynomial based on the screening gate coefficients. Calculate the ciphertext gating coefficient.

[0018] Preferably, after completing the first round of federated learning training, each edge monitoring terminal calculates the model validation loss based on its local training sample set and local model, and sends the model validation loss and the current training round to the cloud server. The cloud server then calculates the model validation loss based on the current training round. and model validation loss Dynamically adjust the physical deviation threshold of the plaintext in this round. And generate a new encryption threshold based on the key.

[0019] Secondly, embodiments of the present invention provide a complex equipment fault federated diagnosis system based on homomorphic encryption, characterized in that it includes several edge monitoring terminals, a trusted third party, and a cloud server; The edge monitoring terminal includes a data acquisition module, a federated training module, a packaging module, and a decryption module. The data acquisition module collects real-time operational data from the target equipment, preprocesses it to build a local training sample set, and transmits this data to the federated training module. The federated training module pre-sets the target equipment's multi-dimensional physical constraint model, baseline physical function relationships, and physical deviation threshold. It uses the local training sample set to perform federated learning training on the local model, calculates the gradient update amount of the model weights, and calculates the consistency residual based on the real-time operational data and baseline physical function relationships. The packaging module uses a public key to encrypt and package the gradient update amount and physical consistency residual into a ciphertext data packet, which is then sent to the cloud server. The physical deviation threshold is also sent to the cloud server. The decryption module decrypts the received global ciphertext model to obtain the updated global model. A trusted third party selects the CKKS fully homomorphic encryption scheme to generate keys, distributes the public key and private key to the entire network, and stores the private key. The cloud server uses a key to encrypt the physical deviation threshold to generate an encrypted threshold. Based on the ciphertext data packet and the encrypted threshold, a global ciphertext model is obtained through homomorphic operation and fed back to the edge monitoring terminal.

[0020] As can be seen from the above technical solution, compared with the prior art, this invention discloses a federated diagnosis method and system for complex equipment faults based on homomorphic encryption, which consists of a trusted third party, an edge monitoring terminal, and a cloud server. The trusted third party is responsible for independent homomorphic key generation and distribution; the edge monitoring terminal, as the data owner, is responsible for collecting data, performing local training and physical consistency verification, and uploading gradients and physical credentials in encrypted form; the cloud server, as the computing core, performs encrypted gating cleaning and model aggregation calculations without obtaining plaintext data and specific physical residual values, achieving complex equipment fault security diagnosis through multi-dimensional physical manifold construction, encrypted gating generation, and privacy-weighted aggregation. The beneficial effects include at least: (1) Improved physical consistency and noise resistance of the model. Compared with traditional aggregation methods that rely solely on statistical distribution of data, this invention introduces "multidimensional physical manifold" as a strong constraint, which can accurately eliminate false samples (such as sensor drift or malicious injection) that violate basic physical laws even though the values ​​are within a reasonable range, ensuring that the diagnostic model focuses on learning real equipment fault characteristics.

[0021] (2) The solution described in this invention achieves blind data cleaning under "zero knowledge". It innovatively designs a homomorphic gating mechanism for the encrypted domain, which enables the cloud server to assess the credibility of the data and complete weighted filtering without decrypting or snooping on the original data of the edge nodes. This perfectly solves the contradiction between "data security protection" and "data quality control" in federated learning.

[0022] (3) This invention has broad applicability and computational efficiency. It adopts "input-state-response" manifold modeling, which makes it directly applicable to a variety of complex equipment such as wind power and gas turbines; combined with SIMD (single instruction multiple data) parallel packing technology with fully homomorphic encryption, it greatly reduces the communication and computation overhead of ciphertext operation and meets the needs of large-scale deployment in industrial scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0024] Figure 1 This is an overall flowchart of the complex equipment fault federated diagnosis method based on homomorphic encryption provided in the embodiments of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention discloses a federated diagnosis method for complex equipment faults based on homomorphic encryption, comprising the following steps: S1: Pre-set multi-dimensional physical constraint models of target equipment in each edge monitoring terminal, input the energy conversion characteristics or operating state manifold of the equipment, and establish the output response. Input incentive variables and control state variables The baseline physical function relationship between them And set the allowable physical deviation threshold. Simultaneously, it is sent to the cloud server; S2: Introduce a trusted third party to select the CKKS fully homomorphic encryption scheme to generate the key, and use the public key... The computing key is distributed to the entire network, and the private key is securely stored. ; S3: In the During each training cycle, each edge monitoring terminal Real-time operational data of the target equipment is collected and preprocessed to construct a local training sample set. Edge monitoring terminals utilize local training sample sets Local samples and the previous global model Train the model and calculate the gradient update of the model weights. ; S4: In a plaintext environment, the edge monitoring terminal substitutes the physical input parameters from the real-time operating data into the baseline physical function relationship, calculates the absolute error between the measured response and the theoretical response, and obtains the physical consistency residual. ; S5: Edge monitoring terminals utilize public keys Encrypt the gradient update amount and the physical consistency residual separately to generate ciphertext gradients. and encrypted physical credentials Package it into an encrypted data packet and upload it to the cloud server; S6: The cloud server uses the key to set a physical deviation threshold. Perform encryption to generate encryption threshold After receiving the encrypted data packet, it processes it without knowing the specific value. First, homomorphic subtraction is used to calculate the ciphertext physical credentials and encryption thresholds for each terminal. deviation ; Secondly, a polynomial approximation algorithm is executed within the ciphertext domain to generate ciphertext gating coefficients based on the deviation. ;like This indicates that the residual exceeds the standard. Approaching encrypted 0; if , indicating that it conforms to the laws of physics, then Approaching encryption 1; Next, homomorphic multiplication is used to multiply the gating coefficients element-wise with the ciphertext gradient to obtain... This suppresses the contribution of poor physical consistency; Finally, homomorphic addition is used to accumulate the filtered ciphertext gradients to obtain the global ciphertext aggregation. ; S8: The cloud server calculates the sum of all ciphertext gating coefficients. To determine the effective number of nodes, the aggregated gradient is normalized using homomorphic division or an interaction protocol. The average gradient is then calculated and the global ciphertext model is updated. ; S9: The updated global encrypted model is broadcast back to each edge monitoring terminal. After decryption, it is used to determine the model convergence status. If it has not converged, it returns to the local computation stage to continue iterating until the training requirements are met.

[0027] Furthermore, after completing the first round of federated learning training, each edge monitoring terminal calculates the model validation loss based on its local training sample set and local model, and sends the model validation loss and the current training round to the cloud server. The cloud server then calculates the model validation loss based on the current training round. and model validation loss The physical deviation threshold of the plaintext in this round is dynamically adjusted, and a new encryption threshold is generated based on the key encryption.

[0028] On the other hand, in one specific embodiment, the specific process of employing a complex equipment fault federated diagnosis method based on homomorphic encryption is as follows: Figure 1 As shown, it specifically includes: S1: Pre-set multi-dimensional physical constraint model of target equipment in each edge monitoring terminal; Specifically, the energy conversion characteristics or operational manifold of the target equipment are entered to establish the output response. Input incentive variables and control state variables The baseline physical function relationship between them And set the allowable physical deviation threshold. Physical deviation threshold It will be sent to the cloud server; S2: Introduce a trusted third party (Key Generation Center) to generate public parameters and a public key using the CKKS fully homomorphic encryption scheme. Private key And the homomorphic computation keys required for homomorphic computation, which include relinearization keys (RelinKeys) and rotation keys (GaloisKeys), where the public key... Used for encryption, private key Used for decryption; public key The homomorphic computation key is distributed to all edge monitoring terminals and cloud servers, along with the private key. Strict confidentiality or multi-party secure computation (MPC) is required for fragmented storage; S3: The cloud server initializes a deep neural network model for fault diagnosis of complex equipment and sets the initial weight parameters. The initial model parameters are then distributed to all edge monitoring terminals participating in the training. S4: In the During each training cycle, each edge monitoring terminal Real-time operational data of the target equipment is collected through sensors, including vibration signals, process status signals, electrical parameters, temperature, and physical input parameters, such as fluid potential energy parameters and flow rate parameters. Each edge monitoring terminal performs noise reduction and normalization processing on the collected data, and aligns sensor data from different sampling frequencies on the time axis to form a standardized input tensor, thus constructing a local training sample set. ; S5: Each edge monitoring terminal utilizes a local training sample set The global model from the previous training cycle Perform forward and backward propagation algorithms to calculate the gradient update amount of the model weights. ; S6: In a local plaintext environment, each edge monitoring terminal substitutes the physical input parameters and control state variables corresponding to the samples in the current batch of the local training sample set into the baseline physical function relationship of the multidimensional physical constraint model to obtain the theoretical response, and extracts the actual output data from the samples as the measured response, calculates the absolute error between the measured response and the theoretical response, and obtains the physical consistency residual. ; When a sample batch contains multiple data points, the physical consistency residuals of each sample in the batch are statistically summarized, and the average, maximum or weighted average of the residuals is taken as the physical consistency residuals corresponding to the samples in the batch. S7: Each edge monitoring terminal uses the CKKS encoding scheme to update high-dimensional gradient values. Pack it into a polynomial form and use the public key. Encryption is performed to generate a ciphertext gradient. ; S8: Each edge monitoring terminal will calculate the physical consistency residuals obtained in S6. Similarly, using the public key Encryption is performed to generate ciphertext physical credentials. Subsequently, the ciphertext gradient and ciphertext physical credentials are packaged into a data packet and sent to the cloud server. S9: Cloud server receives data from... Data packets uploaded by each edge monitoring terminal, and the physical deviation threshold based on the key received from the broadcast. Perform encryption to generate encryption threshold At this point, the cloud server only holds the encrypted data and cannot know the specific gradient values ​​of each edge monitoring terminal, nor can it know the specific physical residual values. S10: The cloud server first uses homomorphic subtraction to calculate the deviation between the ciphertext physical credential and the current encryption threshold to obtain the ciphertext deviation; S11: To eliminate dimensional differences between different physical target equipment (such as different types of sensors), the system performs a ciphertext domain normalization operation:

[0029] in, The encrypted data is the standard deviation of historical residuals and is updated periodically by a trusted third party based on historical data statistics. It is a tiny constant used to prevent division by zero errors; computational implementation: the normalization operation is approximated by homomorphic division using the CKKS algorithm, and the system uses Newton's iteration method to iterate multiple times in the ciphertext domain; S12: The cloud server first utilizes Normalization bias Perform coarse screening calculations to obtain the output coarse screening gate coefficient. It is used to quickly distinguish between clearly valid, clearly invalid, and data in the transition zone; like It is directly judged as invalid data, and the ciphertext gating coefficient is set to zero. Set to 0 to skip subsequent calculations; like This value is directly used as the final ciphertext gating coefficient. ; If the data is in an intermediate state, that is... Then it enters the fine screening stage; application of transition zone data Calculate the screening gate coefficient Call the pre-stored correction polynomial Generate the final ciphertext gating coefficients. ; in, , , In order, they are the third-order Chebyshev polynomial, the seventh-order Chebyshev polynomial, and the second-order correction polynomial. S13: The cloud server performs a homomorphic multiplication operation, and the ciphertext gating coefficients generated in S13 are... Ciphertext gradient uploaded with S8 Element-wise multiplication yields the filtered ciphertext gradient. This step automatically suppresses the contribution of terminals with poor physical consistency to the global model. S14: The cloud server performs a homomorphic addition operation to accumulate the ciphertext gradients filtered by all edge monitoring terminals, obtaining the global ciphertext aggregation value. ; S15: The cloud server synchronously calculates the sum of all ciphertext gating coefficients. The number of nodes effectively participating in the aggregation is used to calculate the average gradient using homomorphic division or an interactive decryption-assisted protocol. And update the global ciphertext model: , For pre-set coefficients; S16: Encryption threshold corresponding to the preset physical deviation threshold It is not fixed; the cloud server adjusts the training round based on the current training time within the edge monitoring terminal. The global model validation loss corresponding to the received global model (local model) Dynamically adjust the physical deviation threshold of the plaintext in this round. The formula is adjusted as follows:

[0030] in, Set as a sensitivity coefficient to control the adjustment range; this round's physical deviation threshold. The encryption threshold is generated by a trusted third party through a secure channel after calculation. The data is then sent to the cloud server; each edge monitoring terminal calculates and verifies the loss locally and sends it to a trusted third party after secure aggregation. S17: The cloud server will update the global ciphertext model. The broadcast is sent back to each edge monitoring terminal. Each terminal decrypts the data to obtain the new plaintext model, and then determines whether the model has converged. If it has not converged, then... Then return to execute S4 until the preset training rounds or accuracy requirements are reached.

[0031] Furthermore, in S1, the output response includes power, efficiency, etc.; the input excitation variables include potential energy, kinetic energy parameters, etc.; the control state variables include opening degree, adjustment amount, etc.; and the physical deviation threshold includes 5% of the rated output index. Furthermore, deep neural network models can use ResNet or Transformer network structures.

[0032] On the other hand, in a specific embodiment, taking a complex equipment fault federated diagnosis system based on homomorphic encryption as an example, the algorithm flow for multidimensional physical manifold construction, ciphertext gating generation, and privacy-weighted aggregation is described. The system consists of a trusted third party, an edge monitoring terminal, and a cloud server. The trusted third party is responsible for independent homomorphic key generation and distribution; the edge monitoring terminal, as the data owner, is responsible for collecting data, performing local training and physical consistency verification, and uploading gradients and physical credentials in encrypted form; the cloud server, as the computing core, performs encrypted gating cleaning and model aggregation calculation without obtaining plaintext data and specific physical residual values.

[0033] Based on the same inventive concept, this invention also provides a complex equipment fault federated diagnosis system based on homomorphic encryption. Since the principles by which these systems solve problems are similar to the aforementioned methods, the implementation of this system can refer to the implementation of the aforementioned methods, and repeated details will not be repeated.

[0034] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0035] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A federated diagnosis method for complex equipment faults based on homomorphic encryption, characterized in that, include: S1: Each edge monitoring terminal has a pre-installed multi-dimensional physical constraint model for the target equipment. The multi-dimensional physical constraint model includes a baseline physical function relationship and a set physical deviation threshold. The physical deviation threshold is sent to the cloud server. S2: A trusted third party selects the CKKS fully homomorphic encryption scheme to generate a key, distributes the public key and private key to the entire network, and stores the private key; S3: Each edge monitoring terminal collects real-time operating data of the target equipment, and after preprocessing, constructs a local training sample set. The local training sample set is used to train the local model through federated learning, calculates the gradient update amount of the model weights, and calculates the consistency residual based on the real-time operating data and the relationship between the benchmark physical function. At the same time, the gradient update amount and the physical consistency residual are encrypted and packaged into a ciphertext data packet and sent to the cloud server using the public key. S4: The cloud server uses a key to encrypt the physical deviation threshold to generate an encrypted threshold. Based on the ciphertext data packet and the encrypted threshold, a global ciphertext model is obtained through homomorphic operation and broadcast back to the edge monitoring terminal. S5: The edge monitoring terminal decrypts the global encrypted model to obtain an updated global model. When the global model converges, training stops; otherwise, it returns to S3 to continue iterating through federated learning.

2. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 1, characterized in that, In S1, the edge monitoring terminal establishes an output response by recording the energy conversion characteristics or operating status manifold of the target equipment. Input incentive variables and control state variables The baseline physical function relationship between them .

3. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 1, characterized in that, In S3, each edge monitoring terminal trains its local model using local samples from its local training sample set during each training cycle and calculates the gradient update amount of the model weights. The updated global model obtained by the edge monitoring terminal during each training cycle is used as the local model.

4. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 1, characterized in that, In S3, each edge monitoring terminal substitutes real-time operating data into the benchmark physical function relationship in a plaintext environment to obtain the theoretical response, and extracts the actual output response from the real-time operating data as the measured response. The absolute difference between the measured response and the theoretical response is calculated to obtain the physical consistency residual.

5. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 1, characterized in that, In S3, the edge monitoring terminal uses a public key to encrypt the gradient update amount to generate a ciphertext gradient, and encrypts the physical consistency residual to generate a ciphertext physical credential.

6. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 5, characterized in that, The specific process by which the cloud server obtains the global ciphertext model through homomorphic operations based on the ciphertext data packet and the encryption threshold is as follows: S41: The cloud server extracts the ciphertext physical credentials from the ciphertext data packet and uses homomorphic subtraction to calculate the deviation between the ciphertext physical credentials and the encryption threshold of each edge monitoring terminal to obtain the ciphertext deviation. S42: The cloud server normalizes the ciphertext deviation and performs a polynomial approximation algorithm within the ciphertext domain based on the normalized ciphertext deviation to generate ciphertext gating coefficients. If the normalized ciphertext deviation is greater than 0, the ciphertext gating coefficients approach the encrypted 0; if the normalized ciphertext deviation is less than or equal to 0, the ciphertext gating coefficients approach the encrypted 1. S43: The cloud server extracts the ciphertext gradient from the ciphertext data packet and uses homomorphic multiplication to multiply the ciphertext gating coefficients with the ciphertext gradient element by element to obtain the filtered ciphertext gradient. S44: The cloud server uses homomorphic addition to accumulate the gradient of the filtered ciphertext to obtain the global ciphertext aggregation. S45: The cloud server calculates the sum of all ciphertext gating coefficients as the number of effective nodes, and uses homomorphic division or interaction protocol to normalize the global ciphertext aggregation amount to obtain the average gradient update amount. S45: The cloud server updates the current global model based on the average gradient update amount to obtain the global encrypted model.

7. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 6, characterized in that, The specific process of S42 is as follows: S421: Use a third-order Chebyshev polynomial to perform coarse screening calculations on the normalized ciphertext bias to obtain the coarse screening gate coefficients. ; S422: Determine the coarse screening gate coefficient If the value is within the valid threshold range, then proceed to fine screening and enter S423. If the coarse screening gate coefficient If the minimum boundary value is less than the effective threshold, then the ciphertext gating coefficient is set to 0; If the coarse screening gate coefficient If the maximum boundary value is greater than the effective threshold, then the coarse screening gate coefficient will be adjusted. Assign the value to the ciphertext gating coefficient; S423: Calculating the screening gating coefficients using a seventh-order Chebyshev polynomial. And call the second-order correction polynomial based on the screening gate coefficients. Calculate the ciphertext gating coefficient.

8. The method for federated diagnosis of complex equipment faults based on homomorphic encryption according to claim 6, characterized in that, After completing the first round of federated learning training, each edge monitoring terminal calculates the model validation loss based on the local training sample set and local model, and sends the model validation loss and the current training round to the cloud server. The cloud server dynamically adjusts the physical deviation threshold for this round based on the current training round and the model validation loss, and generates a new encryption threshold based on the key encryption.

9. A complex equipment fault federated diagnosis system based on homomorphic encryption, characterized in that, This includes several edge monitoring terminals, trusted third parties, and cloud servers; The edge monitoring terminal includes a data acquisition module, a federated training module, a packaging module, and a decryption module; The data acquisition module collects real-time operational data of the target equipment, and after preprocessing, constructs a local training sample set, which is then combined with the real-time operational data and transmitted to the federated training module. The federated training module pre-configures a multi-dimensional physical constraint model of the target equipment, including a baseline physical function relationship and a physical deviation threshold. It uses the local training sample set to perform federated learning training on the local model, calculates the gradient update amount of the model weights, and calculates the consistency residual based on real-time running data and the baseline physical function relationship. The packaging module uses the public key to encrypt and package the gradient update amount and the physical consistency residual into a ciphertext data packet and sends it to the cloud server, and also sends the physical deviation threshold to the cloud server. The decryption module decrypts the received global ciphertext model to obtain the updated global model; A trusted third party selects the CKKS fully homomorphic encryption scheme to generate keys, distributes the public key and private key to the entire network, and stores the private key. The cloud server uses a key to encrypt the physical deviation threshold to generate an encrypted threshold. Based on the ciphertext data packet and the encrypted threshold, a global ciphertext model is obtained through homomorphic operation and fed back to the edge monitoring terminal.