A fusion device state monitoring method, system, device and storage medium

Abstract

The application discloses a fusion device state monitoring method, system, equipment and storage medium, and the method comprises the steps of obtaining historical key operation parameters of a fusion device, and constructing a state prediction model based on an improved time series neural base extension analysis N-BEATS model according to the historical key operation parameters; obtaining current key operation parameters of the fusion device, inputting the current key operation parameters into the state prediction model, and obtaining parameter prediction values; and determining the operation state of the fusion device according to the parameter prediction values, so as to issue a warning when the operation state is abnormal. The application can effectively capture complex nonlinear time series relationships, and improves the accuracy and efficiency of fusion device state prediction.

Description

Technical Field

[0001] This invention relates to the field of fusion device monitoring technology, and in particular to a method, system, device and storage medium for fusion device status monitoring based on an improved N-BEATS (Neural Basis Expansion Analysis for Time Series) model. Background Technology

[0002] Currently, the safety monitoring of EAST (Experimental Advanced Superconducting Tokamak) is usually carried out by the EAST safety interlock system, which monitors the status of sub-control systems such as vacuum and power supply. If the status is in a fault state, the safety interlock system will take protective action.

[0003] Current technologies for monitoring the condition of fusion devices rely solely on real-time detection of subsystem status (normal or abnormal) to execute corresponding protective actions. This approach cannot predict potential failure risks or assess system stability in a timely manner. Traditional prediction methods largely depend on physical models for prediction and risk assessment; however, physical models are often limited by computational complexity, real-time performance, and accuracy, and they perform poorly in fitting nonlinear relationships, failing to meet the requirements for efficient and safe operation of the device. With the development of deep learning technology, data-driven prediction models are widely used in various prediction tasks, especially in handling complex and nonlinear relationships, where deep learning has demonstrated enormous potential. However, existing deep learning methods typically suffer from certain lag and insufficient accuracy in time series prediction.

[0004] Therefore, there is an urgent need for a method for monitoring the state of fusion devices that can handle time series predictions while possessing high accuracy and real-time performance. Summary of the Invention

[0005] To address the above technical problems, this invention provides a method, system, device, and storage medium for monitoring the state of a fusion device, which can effectively capture complex nonlinear time-series relationships and improve the accuracy and efficiency of fusion device state prediction.

[0006] This invention provides a method for monitoring the status of a fusion device, comprising:

[0007] The historical key operating parameters of the fusion device are obtained, and a state prediction model is constructed based on the improved time-series neural base extended analysis N-BEATS model according to the historical key operating parameters.

[0008] Obtain the current key operating parameters of the fusion device, input the current key operating parameters into the state prediction model, and obtain the parameter prediction values;

[0009] Based on the predicted values ​​of the parameters, the operating status of the fusion device is determined, and an early warning is issued when the operating status is abnormal.

[0010] As an improvement to the above scheme, the N-BEATS model includes several stacked modules, each stacked module is composed of several basic modules connected in series, and each basic module includes several fully connected layers; wherein, the input of the first basic module includes a sequence of historical key operating parameters, the output of the basic module includes the predicted value of the key operating parameters and the reconstructed value of the input sequence, and the input of the remaining basic modules includes the residual between the input sequence of the previous basic module and the reconstructed value of the input sequence.

[0011] As an improvement to the above scheme, the N-BEATS model is obtained by introducing a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module into the basic N-BEATS model.

[0012] As an improvement to the above scheme, the multi-scale feature extraction module is set in the basic module, including three parallel one-dimensional convolutional layers and one fully connected layer; wherein, the convolutional layer is used to capture the short-term, medium-term and long-term dependencies of the input target sequence, and the fully connected layer is used to integrate the convolutional features and generate a comprehensive multi-scale feature representation.

[0013] The modern temporal convolutional network is introduced into the model by replacing the original temporal convolutional network in the basic N-BEATS model, and is used to process external features that are synchronously input into the model with key operating parameters.

[0014] The convolutional attention module is set after each stacked module, and the output of the previous stacked module passes through a convolutional attention module before entering the next stacked module.

[0015] As an improvement to the above scheme, after obtaining the historical key operating parameters of the fusion device, the method further includes:

[0016] Based on the historical key operating parameters, the maximum mutual information coefficient of the parameters is calculated, and a set number of parameters are selected in descending order of the maximum mutual information coefficient as the training dataset of the N-BEATS model.

[0017] As an improvement to the above scheme, the step of constructing a state prediction model based on the improved time-series neural base extension analysis N-BEATS model according to the historical key operating parameters includes:

[0018] The model parameters are initialized based on the improved N-BEATS model;

[0019] Using the aforementioned historical key operating parameters as the training dataset, the training dataset is input into the N-BEATS model in batches;

[0020] Forward propagation is performed on each batch of training data to obtain parameter prediction results for future time moments;

[0021] Calculate the loss function based on the difference between the predicted and actual values ​​of the parameters;

[0022] Based on the loss function, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the gradient is passed back to each layer of the model network.

[0023] Based on the gradient, with the goal of minimizing the loss value, the gradient descent optimization algorithm is used to update the model parameters and iteratively optimize until the preset termination condition is met. The optimized N-BEATS model is then output as a state prediction model.

[0024] As an improvement to the above scheme, determining the operating status of the fusion device based on the predicted parameter values ​​includes:

[0025] Determine whether the predicted value of the parameter is greater than a preset safety threshold;

[0026] If all the predicted values ​​of the parameters are less than the corresponding safety threshold, then the fusion device is determined to be operating normally.

[0027] If the predicted value of the parameter is greater than the corresponding safety threshold, the fusion device is determined to be in an abnormal operating state, and an early warning signal is issued.

[0028] The present invention also provides a fusion device status monitoring system, comprising:

[0029] The model building module is used to obtain the historical key operating parameters of the fusion device, and to build a state prediction model based on the improved time series neural base extended analysis N-BEATS model according to the historical key operating parameters.

[0030] The parameter prediction module is used to obtain the current key operating parameters of the fusion device, input the current key operating parameters into the state prediction model, and obtain the parameter prediction values.

[0031] The status monitoring module is used to determine the operating status of the fusion device based on the predicted values ​​of the parameters, and to issue an early warning when the operating status is abnormal.

[0032] The present invention also provides a computer device, including a processor and a memory, wherein the memory stores a computer program and the computer program is configured to be executed by the processor, wherein the processor executes the computer program to implement the fusion device state monitoring method described in any of the preceding claims.

[0033] The present invention also provides a computer-readable storage medium storing a computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the fusion device status monitoring method described above.

[0034] Compared to existing technologies, the beneficial effects of the fusion device state monitoring method, system, device, and storage medium provided by this invention are as follows: By constructing a state prediction model based on an improved N-BEATS model for state prediction, it can handle complex and nonlinear parameter relationships, achieving accurate and stable real-time state monitoring of the device; by introducing a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module into the N-BEATS model, the internal temporal modeling and complex data processing capabilities are enhanced, while providing a multi-scale perspective. The synergistic effect of the three modules improves the model's performance in complex tasks, effectively capturing complex nonlinear temporal relationships and improving prediction accuracy; through an anomaly detection mechanism, potential safety risks can be identified in real time and timely warnings can be issued, enhancing the system's safety monitoring capabilities. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a fusion device status monitoring method provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the structure of an improved N-BEATS model provided in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the MIC values ​​between various parameters provided in the embodiments of the present invention;

[0038] Figure 4 This is a schematic diagram of the pressure and MIC values ​​of various parameters provided in an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram of the model prediction results provided in an embodiment of the present invention;

[0040] Figure 6 This is a schematic diagram of the structure of a fusion device status monitoring system provided in an embodiment of the present invention;

[0041] Figure 7 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0042] 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.

[0043] Please see Figure 1 , Figure 1 This is a flowchart illustrating a fusion device status monitoring method provided in an embodiment of the present invention. The fusion device status monitoring method includes:

[0044] S1: Obtain the historical key operating parameters of the fusion device, and construct a state prediction model based on the improved time-series neural base extended analysis N-BEATS model according to the historical key operating parameters.

[0045] S2: Obtain the current operating parameters of the fusion device, input the current key operating parameters into the state prediction model, and obtain the parameter prediction values;

[0046] S3: Determine the operating status of the fusion device based on the predicted values ​​of the parameters, and issue an early warning when the operating status is abnormal.

[0047] Specifically, addressing the intelligent monitoring requirements of the EAST fusion device's interlocking system, a neural network model is first determined based on functional requirements, including stability, adaptability, complexity, and interpretability. Considering these requirements, this embodiment selects a fully connected N-BEATS model as the original model and constructs a neural network model accordingly. Simultaneously, to improve the model's ability to handle complex data, this embodiment optimizes the N-BEATS model, constructing an improved N-BEATS model for state prediction, thereby improving the accuracy and efficiency of model prediction. Furthermore, existing EAST fusion device interlocking systems typically only monitor the status of a dozen or so sub-control systems, such as vacuum and power supply. To better understand the device's status and anticipate potential failure risks, this embodiment filters and adds numerous process component parameters during operation, such as the temperature, external vacuum pressure, and coil current of components like superconducting leads and buslines, for real-time status monitoring and risk assessment.

[0048] The N-BEATS model is a deep learning model for time series forecasting. Compared to the three mainstream time forecasting frameworks—Recurrent Neural Networks (RNNs), Transformers, and Convolutional Neural Networks (CNNs)—N-BEATS pioneers a new backbone network, using only fully connected networks for time series forecasting. A fully connected network (FCN) is a basic neural network architecture where all nodes in each layer are connected to all nodes in the layers before and after it. Due to its simple, intuitive, and flexible architecture, the FCN exhibits good versatility and wide applicability, making it particularly suitable for handling various data types without sequential dependencies. It possesses powerful expressive capabilities, approximating various complex nonlinear functions, and is suitable for different types of tasks, from classification and regression to time series forecasting. Furthermore, the FCN is easy to parallelize and hardware accelerate, and its optimization methods are mature, resulting in good performance in practical applications. In this invention, by combining the fully connected network with other network modules, it can adapt to more complex scenarios and achieve good performance.

[0049] The basic N-BEATS architecture consists of a series of stacked modules, each focused on modeling a specific component of the time series, such as trend or seasonality. Each stacked module continuously improves prediction accuracy by optimizing the residuals left by the previous stacked module. The input to the N-BEATS model is a historical window of the time series, and the model output is the predicted value for the specified prediction interval.

[0050] As one optional embodiment, the N-BEATS model includes several stacked modules, each stacked module is composed of several basic modules connected in series, and each basic module includes several fully connected layers; wherein, the input of the first basic module includes a sequence of historical key operating parameters, the output of the basic module includes the predicted values ​​of the key operating parameters and the reconstructed values ​​of the input sequence, and the input of the remaining basic modules includes the residual between the input sequence of the previous basic module and the reconstructed values ​​of the input sequence.

[0051] Specifically, the improved N-BEATS model is based on a fully connected network and consists of multiple stacks (i.e., stacked modules). Each stack is composed of multiple blocks (i.e., base modules) connected in series, and each block includes multiple fully connected layers. The first block directly receives the raw input sequence consisting of historical key operating parameters and generates two outputs: the predicted values ​​(forecast) of the operating parameters and the reconstructed values ​​(backcast) of the block's input. Each subsequent block uses the residuals of the previous block as input, progressively fitting components not yet captured by the preceding blocks. The residuals are obtained by subtracting the reconstructed output (backcast) from the block's input sequence. Thus, the model decomposes time-series information, with blocks in different stacks fitting different components of the time series. Finally, the output of the N-BEATS model is obtained by summing the predicted values ​​from all stacks.

[0052] As one of the optional embodiments, the N-BEATS model is obtained by introducing a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module into the basic N-BEATS model.

[0053] As one optional embodiment, the multi-scale feature extraction module is set in the basic module and includes three parallel one-dimensional convolutional layers and one fully connected layer; wherein, the convolutional layers are used to capture the short-term, medium-term and long-term dependencies of the input target sequence, and the fully connected layer is used to integrate the convolutional features and generate a comprehensive multi-scale feature representation.

[0054] The modern temporal convolutional network is introduced into the model by replacing the original temporal convolutional network in the basic N-BEATS model, and is used to process external features that are synchronously input into the model with key operating parameters.

[0055] The convolutional attention module is set after each stacked module, and the output of the previous stacked module passes through a convolutional attention module before entering the next stacked module.

[0056] Specifically, the original N-BEATS model only accepts a single time series as input and cannot utilize other external features such as dates, holidays, or attribute information to assist the model's predictions. To overcome this limitation, the original development team of the N-BEATS model added external features to the original N-BEATS model and encoded these external features using TCN (Temporal Convolutional Networks), enabling the model to improve its prediction performance with the help of auxiliary information.

[0057] Compared to existing technologies, this invention addresses two main issues. First, the N-BEATS model, which originally generates future predictions directly from historical windows, may lack sufficient ability to capture the importance of specific time periods or features. Second, traditional TCNs, compared to the rapidly emerging Transformer and Multilayer Perceptron (MLP) models in the time series domain, suffer from poor convolutional performance. Therefore, for N-BEATS models incorporating external features, this invention modernizes the TCN and modifies it for time series analysis, giving it a larger effective receptive field and better unleashing the potential of convolution in time series analysis, effectively optimizing the limitations of traditional TCNs. Specifically, this invention replaces the TCN in the existing N-BEATS model with ModernTCN (Modern Temporal Convolutional Network) to encode the external features input to the N-BEATS model. Compared to traditional TCNs, ModernTCN aims to more effectively capture long-distance dependencies and complex patterns in external features through large convolutional kernels and modern optimization techniques (such as reparameterization), providing better performance and efficiency and enhancing the model's ability to handle complex covariates.

[0058] Furthermore, a new block type was constructed by introducing MSRR (Multi-Scale Residual Refinement), a multi-scale feature extraction module. Specifically, within the block, three parallel one-dimensional convolutional layers are used to capture the short-term, medium-term, and long-term dependencies of the input sequence, with each layer having a kernel of different sizes. The convolutional features are then averaged over time and integrated through a fully connected layer to generate a comprehensive multi-scale feature representation. Further processing is then performed using fully connected layers and basis functions to ultimately output the corresponding predicted and reconstructed values.

[0059] This embodiment introduces MSRR to replace the simple fully connected layers in the original model, enhancing internal temporal modeling and providing a multi-scale perspective, thus improving the model's ability to perceive dynamic changes at different time scales. The original N-BEATS model primarily processes information through fully connected layers, while MSRR explicitly introduces convolutional operations, enabling it to more effectively capture local patterns within the input sequence and complex dynamics across different time scales. This is more efficient and expressive than existing methods relying solely on fully connected layers. Furthermore, real-world time series often contain multiple periodic or different time-scale patterns; introducing MSRR provides explicit multi-scale analysis capabilities through parallel receptive fields of varying sizes.

[0060] Furthermore, a Convolutional Block Attention Module (CBAM) is introduced into the N-BEATS model. Specifically, a convolutional attention module is added after the output of each stack, enhancing the information interaction between stacked modules and improving the model's interpretability and performance. CBAM employs a dual attention mechanism, including channel attention and spatial attention, enabling it to comprehensively capture key information from features. Compared to existing technologies, the original N-BEATS model achieves information transfer between stacks through residual connections, which is mainly accumulation. In this embodiment, the stack-level attention mechanism allows the model to learn the importance of different stack outputs, achieving more flexible and dynamic information integration and capturing more complex dependencies than simple accumulation. Moreover, attention weights can provide insights, highlighting the contribution of different stacks to the final prediction, thus improving interpretability. Simultaneously, the CBAM attention mechanism also improves model performance in complex tasks.

[0061] Please see Figure 2 This is a schematic diagram of an improved N-BEATS model provided in an embodiment of the present invention, illustrating a specific example. The improved N-BEATS model includes S stacked modules: Stack1, Stack2, ..., StackS; each of the first S-1 stacked modules is connected to a CBAM module. The model input is ( (X), where The time series consists of running parameters, with X representing external features. After passing through several stacked modules, the total predicted value (Global Forecast), i.e., the model output, is obtained. Each stacked module consists of N basic modules: Block1, Block2, ..., BlockN. The stack input is ( , ), where s is the stack module number, and s-1 represents the previous stack module; the input sequence passes through a series of basic modules in the stack module to obtain the stack residual, which is then fed into the next stack module. Each basic module includes three parallel one-dimensional convolutional layers (Conv1d) with different kernel sizes, namely 12, 24, and 48. The block input of the basic module is ( , ), where b is the base module number, and b-1 represents the previous base module. The input of the base module is processed by three parallel one-dimensional convolutional layers for multi-scale convolution, and then enters a fully connected layer (Multi-Scale Fusion, FC) for feature fusion to obtain a hidden representation; subsequently, this hidden representation enters two parallel fully connected layers (FC), with the left FC outputting... This is used to generate predicted values, while the FC output on the right... This is used to generate reconstructed values; finally, , The input basis function mapping layer generates both predicted and reconstructed values ​​(backcasts). The residual between the input sequence of the current basis module and the reconstructed value is input into the next basis module. The predicted values ​​from each basis module are summed to form the stacked predicted value (Stack forecast). The summation of the predicted values ​​from all stacked modules generates the global forecast, which is the model's output. The basis function mapping layer incorporates modern temporal convolutional networks.

[0062] As one optional embodiment, after obtaining the historical key operating parameters of the fusion device, the method further includes:

[0063] Based on the historical key operating parameters, the maximum mutual information coefficient of the parameters is calculated, and a set number of parameters are selected in descending order of the maximum mutual information coefficient as the training dataset of the N-BEATS model.

[0064] Specifically, the training dataset used for model training originates from the actual dynamic parameters of a fusion device during operation. However, complex fusion device systems involve dozens or even thousands of parameter values, including but not limited to magnetic field strength, discharge mode, wall temperature, cooling flow rate, heating power, vacuum level, electron density, plasma displacement, neutral particle density, current density, temperature, and pressure. Therefore, after collecting the device's operating parameters, feature filtering is necessary to enable the model to focus more on important information and reduce the impact of redundant variables on model performance. This invention employs the maximum mutual information coefficient for key parameter filtering. After determining the target parameter to be predicted, the maximum mutual information coefficient between the target parameter and the remaining parameters is calculated. The top three parameters with the highest mutual information coefficients are then selected in descending order to construct the training dataset, thereby improving the training efficiency of the improved model, avoiding overfitting, and simultaneously increasing prediction accuracy.

[0065] The MIC (Maximal Information Coefficient) is a metric used to measure the strength of the association between two variables. It can detect both linear and nonlinear relationships and has higher accuracy than mutual information (MI). The calculation steps for MIC are as follows:

[0066] Let the sample size be n, and the two variables be X and Y. Given two positive integers... , Perform a scatter plot XY consisting of X and Y. OK Column gridding, dividing it into × The grid is then divided, and the joint probability and marginal probability are estimated according to this division, and then the mutual information is calculated. :

[0067]

[0068] in, Let X be the joint probability density of Y. and Let X be the marginal probability density of Y.

[0069] Then, take the different , Maximum mutual information:

[0070]

[0071] in, For maximum mutual information.

[0072] Furthermore, the maximum mutual information is normalized using the following formula:

[0073]

[0074] in, This represents the maximum mutual information after normalization.

[0075] Furthermore, the global maximum value among the maximum mutual information at different scales is selected as the MIC value. :

[0076]

[0077] in, This represents the maximum number of grid divisions, used to limit the number of grids and avoid overcomplicating high-dimensional data. The MIC value is generally normalized to [0, 1], and the closer it is to 1, the stronger the relationship between the two variables.

[0078] As one optional embodiment, the step of constructing a state prediction model based on the improved time-series neural base extension analysis N-BEATS model according to the historical key operating parameters includes:

[0079] The model parameters are initialized based on the improved N-BEATS model;

[0080] Using the aforementioned historical key operating parameters as the training dataset, the training dataset is input into the N-BEATS model in batches;

[0081] Forward propagation is performed on each batch of training data to obtain parameter prediction results for future time moments;

[0082] Calculate the loss function based on the difference between the predicted and actual values ​​of the parameters;

[0083] Based on the loss function, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the gradient is passed back to each layer of the model network.

[0084] Based on the gradient, with the goal of minimizing the loss value, the gradient descent optimization algorithm is used to update the model parameters and iteratively optimize until the preset termination condition is met. The optimized N-BEATS model is then output as a state prediction model.

[0085] Specifically, after obtaining the training dataset, the improved N-BEATS model is trained. First, network parameters are initialized, including the weight matrix, biases, number of training iterations, batch size, and learning rate. The weight matrix and biases are initialized using a Gaussian distribution or Xavier initialization to avoid gradient explosion or vanishing problems. Then, mini-batch data is obtained from the training dataset and fed into the improved N-BEATS model in batches to improve computational efficiency and reduce computational resource consumption. Next, forward propagation is performed on each batch of data, processing the input data through each layer of the model network until the output layer, obtaining the model's prediction results. The forward propagation results predict process parameters for future timeframes. Based on the difference between the predicted and actual values, a loss function (such as mean squared error (MSE) or mean absolute error (MAE)) is calculated to evaluate the model's current predictive ability. Then, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and this gradient information is passed back to each layer of the model network. This backpropagation process allows the network to understand which parameters contribute most to the prediction error, thus guiding optimization. Then, gradient updates are performed. Based on the calculated gradients, the model parameters are updated using a gradient descent optimization algorithm (such as Adaptive Moment Estimation Adam or Stochastic Gradient Descent SGD) to minimize the loss. This process is repeated iteratively until the model achieves the expected fit on the training dataset.

[0086] After thorough training, the improved N-BEATS model will be validated on a test set to evaluate its generalization ability and accuracy. Finally, the trained N-BEATS model will be integrated into the central control system as a state prediction model to process various data from the fusion device in real time and perform parameter predictions.

[0087] As one optional embodiment, determining the operating status of the fusion device based on the predicted parameter values ​​includes:

[0088] Determine whether the predicted value of the parameter is greater than a preset safety threshold;

[0089] If all the predicted values ​​of the parameters are less than the corresponding safety threshold, then the fusion device is determined to be operating normally.

[0090] If the predicted value of the parameter is greater than the corresponding safety threshold, the fusion device is determined to be in an abnormal operating state, and an early warning signal is issued.

[0091] Specifically, after the state prediction model has been fully trained and validated, it is integrated into the central control system of EAST in real time. The system continuously acquires real-time data on operating parameters (such as the temperature at the tip of the high-temperature current lead, external vacuum pressure, and the liquid levels in the five liquid nitrogen tanks) from the sensor network of the EAST fusion device. This data, after preprocessing, is input into the state prediction model. Based on the real-time acquired operating parameters, the state prediction model calculates predicted values ​​for various parameters over future time periods. These predicted values ​​provide a basis for subsequent operational decisions and help engineers understand the potential operating states of the system. Furthermore, based on the predicted values ​​and set safety thresholds, the system can identify potential risks in real time. For example, when predicted parameters such as temperature or pressure exceed the corresponding safety thresholds, the system will issue a warning signal, indicating possible faults or dangers and reminding operators to take appropriate measures in a timely manner. Further, when the system issues a warning signal, the control system will adjust the operating state of the equipment through automated mechanisms to reduce potential risks. During this process, the output of the state prediction model is compared with real-time data to ensure that the control system maintains stable and efficient operation in a constantly changing environment. Through this process, the embodiments of the present invention not only provide high-precision parameter prediction, but also enhance the safety and stability of the system through real-time monitoring and intelligent early warning mechanisms, which helps to improve the safe operation level of the EAST device.

[0092] In a specific example, the pressure parameters of the cooling subsystem in the EAST fusion device were selected for prediction. The improved N-BEATS model provided in this invention was chosen as the state prediction model, and the training data came from real experimental data from 2018 to 2023, with each training data point spaced one hour apart. Key influencing factors corresponding to the pressure parameter prediction were screened using MIC (Micro-Induced Prediction). Please refer to [link / reference]. Figure 3 This displays the MIC values ​​for various parameters, including: Magnetic Field, Discharge Mode, Wall Temperature, Cooling Flow Rate, Heating Power, Vacuum Degree, Electron Density, Plasma Position, Neutral Particle Density, Current Density, Temperature, and Pressure. The last line reveals the correlation between the pressure parameter and the other parameters, specifically as follows: Figure 4 As shown, the three parameters with the largest mutual information coefficients are selected: temperature, discharge mode, and current density. These three parameters are then used as the key influencing factors corresponding to the pressure parameter prediction, i.e., the key operating parameters.

[0093] Please see Figure 5 This diagram illustrates the training and testing results of the state prediction model. The solid blue line represents the actual measured pressure value; the dashed orange line represents the model's prediction results (the last 7 days of the dataset); and the vertical black line distinguishes between historical data and the future prediction phase. During the prediction period (to the right of the black line), the pressure trend predicted by the model is largely consistent with the actual data. The model in this embodiment of the invention captures the periodic changes and oscillations of pressure well; simultaneously, the fluctuation ranges of the predicted and actual values ​​are close, demonstrating a good fitting effect. Based on the predicted values ​​and the set safety threshold, the system can identify potential risks in real time. When the predicted pressure parameter exceeds the set threshold, the system will issue a warning signal, indicating possible faults or dangers, and reminding operators to take appropriate measures in a timely manner.

[0094] This invention, through an improved N-BEATS model architecture, enables real-time prediction of multivariate process parameters in the EAST system, enhancing the real-time performance of status monitoring. It effectively captures complex nonlinear time-series relationships and improves the modeling accuracy of the dynamic behavior of the EAST device. By introducing external features (such as environmental changes and operating parameters), the model's adaptability to system behavior under the influence of multiple factors is enhanced, further improving prediction accuracy. This invention can monitor the EAST device status in real time and issue early warnings based on prediction results, helping operators take necessary measures to avoid faults or accidents. The accurate prediction and intelligent early warning mechanism improves the safety and reliability of the EAST device, effectively avoiding potential faults and ensuring stable and safe operation.

[0095] Furthermore, the model in this embodiment of the invention exhibits strong noise resistance, reducing noise interference from various sources that may affect the EAST device's data, such as sensor measurement errors and sudden abnormal states of the system. This prevents predictions from being affected by anomalous data, thus improving model stability. The model is highly interpretable, capable of decomposing time series data into trend, seasonal, and exogenous variables, which helps in understanding the contribution of each component to the prediction. The decomposition of trend and seasonality helps in understanding long-term changes and periodic fluctuations in the system, further improving the accuracy of state predictions. Through backpropagation and gradient descent optimization algorithms, the network weights and biases can be automatically adjusted to adapt to different data patterns and task requirements, enabling the processing of complex and variable data. The model of this invention also possesses advantages such as parallel computing and ease of integration, ensuring efficiency.

[0096] Accordingly, the present invention also provides a fusion device status monitoring system, which can implement all the processes of the fusion device status monitoring method in the above embodiments.

[0097] Please see Figure 6 , Figure 6 This is a schematic diagram of a fusion device status monitoring system provided in an embodiment of the present invention. The fusion device status monitoring system includes:

[0098] The model building module 601 is used to obtain the historical key operating parameters of the fusion device, and to build a state prediction model based on the improved time series neural base extended analysis N-BEATS model according to the historical key operating parameters.

[0099] The parameter prediction module 602 is used to obtain the current key operating parameters of the fusion device, input the current key operating parameters into the state prediction model, and obtain the parameter prediction values.

[0100] The status monitoring module 603 is used to determine the operating status of the fusion device based on the predicted values ​​of the parameters, so as to issue an early warning when the operating status is abnormal.

[0101] Preferably, the N-BEATS model includes several stacked modules, each stacked module is composed of several basic modules connected in series, and each basic module includes several fully connected layers; wherein, the input of the first basic module includes a sequence of historical key operating parameters, the output of the basic module includes the predicted value of the key operating parameters and the reconstructed value of the input sequence, and the input of the remaining basic modules includes the residual between the input sequence of the previous basic module and the reconstructed value of the input sequence.

[0102] Preferably, the N-BEATS model is obtained by introducing a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module into the basic N-BEATS model.

[0103] Preferably, the multi-scale feature extraction module is set in the basic module and includes three parallel one-dimensional convolutional layers and one fully connected layer; wherein, the convolutional layers are used to capture the short-term, medium-term and long-term dependencies of the input target sequence, and the fully connected layer is used to integrate the convolutional features and generate a comprehensive multi-scale feature representation;

[0104] The modern temporal convolutional network is introduced into the model by replacing the original temporal convolutional network in the basic N-BEATS model, and is used to process external features that are synchronously input into the model with key operating parameters.

[0105] The convolutional attention module is set after each stacked module, and the output of the previous stacked module passes through a convolutional attention module before entering the next stacked module.

[0106] Preferably, the fusion device status monitoring system is further used for:

[0107] Based on the historical key operating parameters, the maximum mutual information coefficient of the parameters is calculated, and a set number of parameters are selected in descending order of the maximum mutual information coefficient as the training dataset of the N-BEATS model.

[0108] Preferably, the step of constructing a state prediction model based on the improved time-series neural base extension analysis (N-BEATS) model according to the historical key operating parameters includes:

[0109] The model parameters are initialized based on the improved N-BEATS model;

[0110] Using the aforementioned historical key operating parameters as the training dataset, the training dataset is input into the N-BEATS model in batches;

[0111] Forward propagation is performed on each batch of training data to obtain parameter prediction results for future time moments;

[0112] Calculate the loss function based on the difference between the predicted and actual values ​​of the parameters;

[0113] Based on the loss function, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the gradient is passed back to each layer of the model network.

[0114] Based on the gradient, with the goal of minimizing the loss value, the gradient descent optimization algorithm is used to update the model parameters and iteratively optimize until the preset termination condition is met. The optimized N-BEATS model is then output as a state prediction model.

[0115] Preferably, determining the operating status of the fusion device based on the predicted parameter values ​​includes:

[0116] Determine whether the predicted value of the parameter is greater than a preset safety threshold;

[0117] If all the predicted values ​​of the parameters are less than the corresponding safety threshold, then the fusion device is determined to be operating normally.

[0118] If the predicted value of the parameter is greater than the corresponding safety threshold, the fusion device is determined to be in an abnormal operating state, and an early warning signal is issued.

[0119] In specific implementation, the working principle, control process and technical effects of the fusion device status monitoring system provided in this embodiment of the invention are the same as those of the fusion device status monitoring method in the above embodiments, and will not be repeated here.

[0120] See Figure 7 , Figure 7 This is a structural block diagram of a computer device provided in an embodiment of the present invention. The computer device includes: a processor 701, a memory 702, and a computer program stored in the memory 702 and executable on the processor 701. When the processor 701 executes the computer program, it implements the steps in the above-described fusion device state monitoring method embodiment. Alternatively, when the processor 701 executes the computer program, it implements the functions of each module / unit in the above-described system embodiments.

[0121] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 702 and executed by the processor 701 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the computer device.

[0122] The computer device may include, but is not limited to, a processor 701 and a memory 702. Those skilled in the art will understand that the schematic diagram is merely an example of a computer device and does not constitute a limitation on the computer device. It may include more or fewer components than illustrated, or combine certain components, or different components. For example, the computer device may also include input / output devices, network access devices, buses, etc.

[0123] The processor 701 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor 701 is the control center of the computer device, connecting various parts of the entire computer device through various interfaces and lines.

[0124] The memory 702 can be used to store the computer programs and / or modules. The processor 701 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 702 and calling the data stored in the memory 702. The memory 702 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 702 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0125] Wherein, if the modules / units integrated into the computer device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor 701, it can implement the steps of the various method embodiments described above. Wherein, the computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0126] This invention also provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the fusion device status monitoring method described in any of the above embodiments.

[0127] This invention provides a method, system, device, and storage medium for monitoring the state of a fusion device. Its advantages include: Firstly, by constructing a state prediction model based on an improved N-BEATS model, it can handle complex and nonlinear parameter relationships, achieving accurate and stable real-time state monitoring of the device. Secondly, by introducing a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module into the N-BEATS model, it enhances the internal temporal modeling and complex data processing capabilities, while providing a multi-scale perspective. The synergistic effect of these three modules improves the model's performance in complex tasks, effectively capturing complex nonlinear temporal relationships and improving prediction accuracy. Thirdly, through an anomaly detection mechanism, it can identify potential safety risks in real time and issue timely warnings, enhancing the system's safety monitoring capabilities.

[0128] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for monitoring the condition of a fusion device, characterized in that, include: Historical key operating parameters of the fusion device are obtained, and a state prediction model is constructed based on an improved time-series neural network extended analysis (N-BEATS) model using these parameters. The N-BEATS model comprises several stacked modules, each consisting of several basic modules connected in series, and each basic module including several fully connected layers. The input of the first basic module includes a sequence of the historical key operating parameters, and the output of the basic module includes predicted values ​​of the key operating parameters and reconstructed values ​​of the input sequence. The inputs of the remaining basic modules include the residuals between the input sequence of the previous basic module and the reconstructed values ​​of that input sequence. The N-BEATS model is improved by introducing multi-scale... The multi-scale feature extraction module is obtained by using a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module. The multi-scale feature extraction module, located within the basic module, includes three parallel one-dimensional convolutional layers and one fully connected layer. The convolutional layers capture short-term, medium-term, and long-term dependencies of the input sequence, while the fully connected layer integrates the convolutional features to generate a comprehensive multi-scale feature representation. The modern temporal convolutional network, introduced by replacing the original temporal convolutional network in the basic N-BEATS model, processes external features that are synchronously input into the model with key operating parameters. The convolutional attention module is located after each stacked module; the output of the previous stacked module passes through one of these convolutional attention modules before entering the next stacked module. Obtain the current key operating parameters of the fusion device, input the current key operating parameters into the state prediction model, and obtain the parameter prediction values; Based on the predicted values ​​of the parameters, the operating status of the fusion device is determined, and an early warning is issued when the operating status is abnormal.

2. The fusion device status monitoring method as described in claim 1, characterized in that, After obtaining the historical key operating parameters of the fusion device, the method further includes: Based on the historical key operating parameters, the maximum mutual information coefficient of the parameters is calculated, and a set number of parameters are selected in descending order of the maximum mutual information coefficient as the training dataset of the N-BEATS model.

3. The fusion device status monitoring method as described in claim 1, characterized in that, The process of constructing a state prediction model based on the improved time-series neural base extension analysis (N-BEATS) model using the historical key operating parameters includes: The model parameters are initialized based on the improved N-BEATS model; Using the aforementioned historical key operating parameters as the training dataset, the training dataset is input into the N-BEATS model in batches; Forward propagation is performed on each batch of training data to obtain parameter prediction results for future time moments; Calculate the loss function based on the difference between the predicted and actual values ​​of the parameters; Based on the loss function, the gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and the gradient is passed back to each layer of the model network. Based on the gradient, with the goal of minimizing the loss value, the gradient descent optimization algorithm is used to update the model parameters and iteratively optimize until the preset termination condition is met. The optimized N-BEATS model is then output as a state prediction model.

4. The fusion device status monitoring method as described in claim 1, characterized in that, Determining the operating status of the fusion device based on the predicted parameter values ​​includes: Determine whether the predicted value of the parameter is greater than a preset safety threshold; If all the predicted values ​​of the parameters are less than the corresponding safety threshold, then the fusion device is determined to be operating normally. If the predicted value of the parameter is greater than the corresponding safety threshold, the fusion device is determined to be in an abnormal operating state, and an early warning signal is issued.

5. A fusion device status monitoring system, characterized in that, include: A model building module is used to acquire historical key operating parameters of the fusion device and, based on these parameters, construct a state prediction model using an improved time-series neural base extension analysis (N-BEATS) model. The N-BEATS model comprises several stacked modules, each consisting of several basic modules connected in series, and each basic module including several fully connected layers. The input of the first basic module includes a sequence of the historical key operating parameters, and the output includes predicted values ​​of the key operating parameters and reconstructed values ​​of the input sequence. The inputs of the remaining basic modules include the residuals between the input sequence of the previous basic module and the reconstructed values ​​of that input sequence. The N-BEATS model is constructed by using the basic N-BEATS model... The model incorporates a multi-scale feature extraction module, a modern temporal convolutional network, and a convolutional attention module. The multi-scale feature extraction module, located within the basic module, comprises three parallel one-dimensional convolutional layers and one fully connected layer. The convolutional layers capture short-term, medium-term, and long-term dependencies of the input sequence, while the fully connected layer integrates the convolutional features to generate a comprehensive multi-scale feature representation. The modern temporal convolutional network, introduced by replacing the original temporal convolutional network in the basic N-BEATS model, processes external features that are synchronously input into the model with key operating parameters. The convolutional attention module is placed after each stacked module; the output of the previous stacked module passes through this convolutional attention module before entering the next stacked module. The parameter prediction module is used to obtain the current key operating parameters of the fusion device, input the current key operating parameters into the state prediction model, and obtain the parameter prediction values. The status monitoring module is used to determine the operating status of the fusion device based on the predicted values ​​of the parameters, and to issue an early warning when the operating status is abnormal.

6. A computer device, characterized in that, The device includes a processor and a memory, the memory storing a computer program configured to be executed by the processor, wherein the processor, when executing the computer program, implements the fusion device state monitoring method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the device containing the computer-readable storage medium executes the computer program, it implements the fusion device status monitoring method as described in any one of claims 1 to 4.

Images