A collaborative fault diagnosis method for fusion reactor dynamic equipment based on multi-agent

Abstract

This invention relates to the fields of fault diagnosis and intelligent manufacturing technology, and particularly to a collaborative fault diagnosis method for fusion reactor dynamic equipment based on multi-agent systems. The technical solution includes the following steps: constructing a multi-agent system, deploying vibration agents, temperature agents, and electromagnetic agents for the fusion reactor dynamic equipment, respectively used to collect vibration data, temperature data, and electromagnetic data; each agent performs signal conditioning on the collected data and extracts common features; based on a federated learning framework, multimodal distributed training is performed, with each agent independently training its model to obtain local gradients; each agent encrypts its local gradients and uploads them to a cloud server for secondary joint model training; the cloud server encrypts and distributes the jointly trained model parameters to each agent. This invention significantly improves the accuracy, comprehensiveness, and security of fault diagnosis for fusion reactor dynamic equipment through multi-agent collaboration, multimodal data fusion, and privacy protection through federated learning.

Description

Technical Field

[0001] This invention relates to the fields of fault diagnosis and intelligent manufacturing technology, and in particular to a collaborative fault diagnosis method for fusion reactor dynamic equipment based on multi-agent systems. Background Technology

[0002] Currently, during the operation of nuclear fusion reactors, moving equipment (such as superconducting magnet systems, cryogenic pump sets, and divertor drive mechanisms) serves as the core device for maintaining plasma confinement and energy conversion under extreme conditions. They are subjected to multiple coupled effects from high-intensity electromagnetic fields, thermal stress shocks, and complex mechanical vibrations. Traditional fault diagnosis methods rely on single sensor data sources for analysis, making it difficult to effectively capture abnormal signs of multi-physics interactions. For example, the early stage of magnet quenching may only manifest as magnetic field distribution distortion and slight local temperature field shifts, while wear of mechanical transmission mechanisms may exhibit high-frequency harmonic characteristics in the vibration spectrum. These weakly correlated signals across modes are easily ignored by conventional single-dimensional analysis methods. To address the problems of existing technologies, this invention provides a collaborative fault diagnosis method for fusion reactor moving equipment based on multi-agent systems. Summary of the Invention

[0003] The purpose of this invention is to address the problem in the background art that traditional fault diagnosis methods rely on a single sensor data source for analysis, making it difficult to effectively capture abnormal signs of multi-physics interaction. The invention proposes a collaborative fault diagnosis method for fusion reactor dynamic equipment based on multiple agents.

[0004] The technical solution of this invention: A collaborative fault diagnosis method for fusion reactor propulsion equipment based on multi-agent systems, comprising the following steps:

[0005] A multi-agent system is constructed, deploying vibration agents, temperature agents, and electromagnetic agents for the fusion reactor propulsion equipment, which are used to collect vibration data, temperature data, and electromagnetic data, respectively.

[0006] Each agent performs signal conditioning on the collected data and extracts common features;

[0007] Based on the federated learning framework, multimodal distributed training is carried out, and each agent independently trains the model to obtain local gradients.

[0008] Each agent encrypts its local gradients and uploads them to the cloud server for secondary joint training of the model;

[0009] The cloud server encrypts and distributes the jointly trained model parameters to each agent for multi-agent collaborative fault diagnosis.

[0010] Optionally, the vibration agent collects vibration data based on the IEPE accelerometer vibration sensor and uses an adaptive resonant wavelet packet decomposition algorithm for signal conditioning, wherein the adaptive noise threshold update formula is:

[0011] ,in, For the current moment The noise threshold The next moment The noise threshold Step size, This represents the gradient value of the mean square error function;

[0012] Denoising signal DS( The formula for calculating ) is:

[0013] ,in, The raw acceleration signal acquired by the vibration sensor. For a moment The vibration signal value, For symbolic functions, Noise threshold It is an exponential function.

[0014] Optionally, the temperature agent collects temperature data based on distributed optical fiber and thermal infrared imaging technology, and constructs a three-dimensional temperature field reconstruction model, the model expression of which is:

[0015] ,in, any point in three-dimensional space The reconstruction temperature value, For the The three-dimensional spatial coordinates of each measurement point For the Temperature coefficient at each measurement point This is the measured temperature value at that location. This is the temperature error value at that location. This represents the total number of points in the temperature field.

[0016] Optionally, the electromagnetic intelligent agent uses three-dimensional vector fluxgate technology to measure the magnetic field strength, and the magnetic flux calculation formula is as follows:

[0017] ,in, for Magnetic flux at time It represents the magnetic flux density. The area is perpendicular to the direction of the magnetic field.

[0018] Optionally, the three-modal data are converted into image signals after Z-score normalization for training. The vibration signal is converted into a time-frequency graph through time-frequency analysis, and the temperature and electromagnetic three-dimensional data are converted into standard images.

[0019] Optionally, during federated learning, an improved differential privacy algorithm is used to encrypt the transmission of local features. Privacy noise satisfying the Laplace probability distribution is added to prevent differential attacks. The feature gradient output distribution expression is:

[0020] ,in, For the characteristic gradient output distribution, For query functions, For privacy noise, For the characteristic data of the vibrating intelligent agent, For the characteristic data of the temperature-sensitive agent, For the feature data of the electromagnetic intelligent agent, the privacy noise follows the Laplace probability distribution:

[0021] ,in, For the Laplace probability distribution, This is the range parameter.

[0022] Optionally, homomorphic encryption algorithms are deployed in the cloud and on each intelligent agent to achieve encrypted transmission of local gradients. The gradient parameter encryption process is as follows:

[0023] ,in, For plain text, It is a ciphertext. For encryption operations, To randomly select a value less than positive integers, Less than random numbers, For public key, ;

[0024] The gradient parameter decryption process is as follows:

[0025] ,in, For decryption operation, For private key, For a specific function, specifically , As the independent variable, For public key, It is a random number.

[0026] Compared with the prior art, this application includes at least one of the following beneficial technical effects:

[0027] By deploying three types of intelligent agents—vibration, temperature, and electromagnetic—to collect and analyze multi-physics field data, we can effectively capture cross-modal weak correlation signals and overcome the shortcomings of traditional single-dimensional analysis methods.

[0028] Based on federated learning, a collaborative analysis encryption mode is designed to encrypt and upload local gradients instead of the original data, thereby preventing the leakage of enterprise device operation data and ensuring data privacy and security.

[0029] By employing adaptive resonant wavelet packet decomposition, three-dimensional temperature field reconstruction, three-dimensional vector fluxgate technology, and improved differential privacy and homomorphic encryption algorithms, the accuracy of signal processing and the security of gradient transmission are enhanced, enabling joint training and dynamic updating of diagnostic models.

[0030] This invention significantly improves the accuracy, comprehensiveness, and safety of fault diagnosis for fusion reactor dynamic equipment through multi-agent collaboration, multi-modal data fusion, and federated learning for privacy protection. Attached Figure Description

[0031] Figure 1 This is a flowchart of a collaborative fault diagnosis method for fusion reactor dynamic equipment based on multi-agent systems. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0033] Example

[0034] See Figure 1 This invention proposes a collaborative fault diagnosis method for fusion reactor propulsion equipment based on multi-agent systems. For fusion reactor propulsion equipment, multiple agents are constructed for data acquisition, analysis, and scheduling. Considering the unique characteristics of fusion reactor propulsion equipment, vibration agents, temperature agents, and electromagnetic agents are deployed. Data acquisition and signal conditioning are performed through these three types of agents, extracting their respective common features. Based on a federated learning design, multimodal distributed training is conducted. Local gradients are encrypted and uploaded for secondary model training, achieving the collaborative diagnosis function of multiple agents.

[0035] The vibration agent uses an IEPE accelerometer for data acquisition, and signal conditioning is based on an adaptive resonant wavelet packet decomposition algorithm. The accelerometer signal is... Let the noise threshold be N, the noise threshold at the current time be N(t), and the noise threshold at the next time be N(t+1). Then the adaptive threshold relationship is:

[0036] ,in, For the current moment The noise threshold The next moment The noise threshold Step size, This represents the gradient value of the mean squared error function.

[0037] After adaptive resonant wavelet decomposition, the denoised signal DS( )for

[0038] ,in, The raw acceleration signal acquired by the vibration sensor. For a moment The vibration signal value, For symbolic functions, Noise threshold It is an exponential function.

[0039] The temperature-sensitive agent, based on distributed fiber optic and thermal infrared imaging technologies, acquires temperature data and constructs a three-dimensional temperature field reconstruction model. This model includes a single-point temperature field in three-dimensional space. The temperature field model reconstructed based on temperature data is then...

[0040] ,in, any point in three-dimensional space The reconstruction temperature value, For the The three-dimensional spatial coordinates of each measurement point For the Temperature coefficient at each measurement point This is the measured temperature value at that location. This is the temperature error value at that location. This represents the total number of points in the temperature field.

[0041] The electromagnetic intelligent agent uses three-dimensional vector fluxgate technology to measure the magnetic field strength of the device in a fusion environment, and obtains the magnetic flux value.

[0042] ,in, Let be the magnetic flux at time t. It represents the magnetic flux density. The area is perpendicular to the direction of the magnetic field.

[0043] After Z-score normalization, the three-modal data were all converted into image signals for training. Specifically, the vibration agent converted the vibration signals into a time-frequency graph using time-frequency analysis. The temperature and electromagnetic agents converted the 3D data into standard images.

[0044] The three agents independently train their models to obtain their respective local gradients. Subsequently, based on a federated learning framework, an improved differential privacy algorithm is used to encrypt and transmit the agents' local features to a cloud server for joint training. To prevent differential attacks, a query protection mechanism is designed for multiple agents by adding noise to make the probabilities of querying the same value similar.

[0045] ,in, For the characteristic gradient output distribution, For query functions, This is for privacy noise. For the characteristic data of the vibrating intelligent agent, For the characteristic data of the temperature-sensitive agent, These are the feature data of the electromagnetic intelligent agent. The privacy noise follows a Laplace probability distribution.

[0046] ,in, For the Laplace probability distribution, This is the range parameter.

[0047] Homomorphic encryption algorithms are deployed in the cloud and on each intelligent agent to achieve encrypted uploading and distribution of local gradients. The gradient parameter encryption process is as follows:

[0048] ,in, For plain text, It is a ciphertext. For encryption operations, To randomly select a value less than positive integers, Less than random numbers, For public key, .

[0049] The gradient parameter decryption process is as follows:

[0050] ,in, For decryption operation, For private key, For a specific function, specifically , As the independent variable, For public key, It is a random number.

[0051] By encrypting the local gradients of multiple agents for communication, joint training is conducted in the cloud, and then encrypted data is distributed to each agent, achieving collaborative training and privacy protection of data. This invention designs three special agents: vibration agents, temperature agents, and electromagnetic agents, to collect and analyze vibration, temperature, and electromagnetic multi-physics field data from fusion reactor propulsion equipment, respectively. Through multimodal data fusion and collaborative diagnosis, it can effectively capture cross-modal weak correlation signals (such as magnetic field distortion and temperature shift in the early stage of magnet quenching, and high-frequency vibration harmonics of mechanical wear) that are easily overlooked by traditional single-dimensional analysis methods. This solves the problem that traditional methods rely on a single sensor data source and are unable to reflect abnormal signs of multi-physics field coupling. This invention adopts a federated learning framework to design a collaborative analysis encryption mode. Each agent only encrypts and uploads local gradient parameters (not raw data) to the cloud for joint training, avoiding the leakage of original equipment operation data and meeting the data privacy protection requirements in industrial scenarios. Through encrypted transmission of gradient parameters and joint training in the cloud, a dynamic update closed loop for the diagnostic model is achieved, solving the problem of model inability to train online and delayed updates caused by data privacy restrictions.

[0052] Among them, the vibration intelligent agent adopts the adaptive resonant wavelet packet decomposition algorithm, and dynamically updates the noise threshold (formula). The nonlinear denoising strategy effectively suppresses noise interference under complex working conditions and improves the accuracy of vibration signal feature extraction.

[0053] In addition, the temperature agent constructs a three-dimensional temperature field model based on distributed optical fiber and thermal infrared imaging technology (formula). This enables a three-dimensional representation of the temperature distribution of the equipment and allows for precise location of local thermal anomalies.

[0054] It is worth noting that the electromagnetic intelligent agent accurately measures the magnetic field strength using three-dimensional vector fluxgate technology, combined with an improved differential privacy algorithm (adding Laplace noise) and a homomorphic encryption algorithm (formula). and This ensures that gradient parameters resist differential attacks and privacy leaks during transmission, achieving secure collaborative training where "data is usable but not visible".

[0055] In summary, this invention significantly improves the accuracy, comprehensiveness, and safety of fault diagnosis for fusion reactor dynamic equipment through core technologies such as multi-agent collaboration, multimodal data fusion, and federated learning for privacy protection.

[0056] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A collaborative fault diagnosis method for fusion reactor propulsion equipment based on multi-agent systems, characterized in that, Includes the following steps: A multi-agent system is constructed, deploying vibration agents, temperature agents, and electromagnetic agents for the fusion reactor's propulsion equipment, respectively for collecting vibration, temperature, and electromagnetic data. The vibration agent collects vibration data based on the IEPE accelerometer vibration sensor and employs an adaptive resonant wavelet packet decomposition algorithm for signal conditioning. The adaptive noise threshold update formula is as follows: N(t+1)=N(t)-τΔN(t), Where N(t) is the noise threshold at the current time t, N(t+1) is the noise threshold at the next time t+1, τ is the step size, and ΔN(t) is the gradient value of the mean square error function. Denoising signal DS(D) v The formula for calculating ) is: Among them, D v D is the raw acceleration signal acquired by the vibration sensor. v (t) represents the vibration signal value at time t, sgn(·) is the sign function, N is the noise threshold, and exp(·) is the exponential function; Each agent performs signal conditioning on the collected data and extracts common features; Based on the federated learning framework, multimodal distributed training is performed, with each agent independently training the model to obtain local gradients. During the federated learning process, an improved differential privacy algorithm is used to encrypt the transmission of local features. Privacy noise satisfying the Laplace probability distribution is added to prevent differential attacks. The feature gradient output distribution expression is as follows: M(D v ,D t ,D m )=F(D v ,D t ,D m )+ε, Where M is the feature gradient output distribution, F is the query function, ε is the privacy noise, and D v For the characteristic data of the vibrational intelligent agent, D t For the characteristic data of the temperature-sensitive agent, D m For the feature data of the electromagnetic intelligent agent, the privacy noise follows the Laplace probability distribution: Where P(ε) is the Laplace probability distribution, and γ is the interval range parameter; Each agent encrypts its local gradients and uploads them to the cloud server for secondary joint training of the model; The cloud server encrypts and distributes the jointly trained model parameters to each agent for multi-agent collaborative fault diagnosis.

2. The method for collaborative fault diagnosis of fusion reactor propulsion equipment based on multi-agent systems according to claim 1, characterized in that, The temperature agent collects temperature data based on distributed optical fiber and thermal infrared imaging technology, and constructs a three-dimensional temperature field reconstruction model. The model expression is as follows: Wherein, T(x,y,z) is the reconstructed temperature value of any point (x,y,z) in the three-dimensional space, (x i ,y i ,z i Let α be the coordinates of the i-th measurement point in three-dimensional space. i Let K(x) be the temperature coefficient at the i-th measurement point. i ,y i ,z i ) represents the measured temperature at that location, ε t Let Q be the temperature error value at that location, and let Q be the total number of points in the temperature field.

3. The method for collaborative fault diagnosis of fusion reactor propulsion equipment based on multi-agent systems according to claim 1, characterized in that, The electromagnetic intelligent agent uses three-dimensional vector fluxgate technology to measure magnetic field strength. The formula for calculating magnetic flux is: in, Let t be the magnetic flux, B be the magnetic induction intensity, and S be the area perpendicular to the magnetic field direction.

4. The method for collaborative fault diagnosis of fusion reactor propulsion equipment based on multi-agent systems according to claim 1, characterized in that, The three-modal data were converted into image signals for training after Z-score normalization. The vibration signal was converted into a time-frequency graph through time-frequency analysis, and the temperature and electromagnetic three-dimensional data were converted into standard images.

5. The method for collaborative fault diagnosis of fusion reactor propulsion equipment based on multi-agent systems according to claim 1, characterized in that, Homomorphic encryption algorithms are deployed in the cloud and on each intelligent agent to achieve encrypted transmission of local gradients. The gradient parameter encryption process is as follows: c=E(m)=g m r n mod(n), Where m is the plaintext, c is the ciphertext, E is the encryption operation, and g is the random selection of a value less than n. 2 The positive integer r is a random number less than n, n is the public key, and mod(·) is the modulo operation; the gradient parameter decryption process is as follows: m=D(c)=L(c γ modern 2 )*μmodn, Where D is the decryption operation, γ is the private key, L is a specific function, specifically L(x) = (x-1) / n, x is the independent variable, n is the public key, and μ is a random number.

Images