A real-time fault detection method for tokamak magnetic probes

Abstract

The present invention relates to the technical field of tokamak magnetic probe fault detection, and in particular to a real-time fault detection method for tokamak magnetic probes. The technical solution includes the following steps: collecting neural network model training data to construct a training database, and using physical parameters related to magnetic probe signals as model inputs; using a neural network model with multiple encoding layers and decoding layers to capture the temporal evolution characteristics of magnetic probe signals, and predicting magnetic probe signals in subsequent time periods based on the input time periods; transmitting the model input to the trained neural network model to predict the magnetic probe signals in real time. The present invention proposes to use an artificial intelligence algorithm to realize fault detection of magnetic probe diagnosis in a tokamak device, which not only improves the calculation accuracy of the model, but also greatly optimizes the model calculation time, making real-time detection of magnetic probe faults feasible and having strong universality.

Description

Technical Field

[0001] The invention relates to the technical field of tokamak magnetic probe fault detection, and in particular to a real-time fault detection method for a tokamak magnetic probe. Background Art

[0002] With the continuous development of magnetic confinement fusion research, the safe and stable operation of the device has attracted much attention. In the magnetic confinement tokamak device, there are strong magnetic fields, which are generated by the current flowing in the external conductor (such as the longitudinal field and poloidal field conductor) or the plasma itself. These magnetic fields interact with the plasma to confine the high-temperature plasma in the vacuum chamber. Therefore, in the tokamak plasma experiment, the measurement of the magnetic field is one of the most basic conditions for the stable operation of the device, and the electromagnetic measurement diagnosis system is a simple and effective tool for measuring the magnetic field and electromagnetic process. Based on the inversion of electromagnetic measurement diagnosis, the state of the high-temperature plasma inside the device can be obtained, and then the stable control of the plasma can be achieved. Among them, the magnetic probe is one of the basic diagnoses in electromagnetic measurement, which has the characteristics of simple measurement method and high measurement accuracy. However, the harsh operating environment of the fusion device may cause the magnetic probe to have measurement failures during long-term operation, thus affecting the stable operation of the device. In order to ensure the stable operation of the device, it is necessary to detect the operating status of the magnetic probe in real time.

[0003] Existing magnetic probe fault diagnosis and detection has the following disadvantages:

[0004] (1) It is difficult to construct an evolutionary model

[0005] The discharge process of a tokamak is an extremely complex process, and it is difficult to accurately simulate the traditional physical evolution. The existing technology generally simplifies the physical process based on certain assumptions to achieve discharge simulation. The simplified physical process can only adapt to part of the discharge process. When the discharge parameters are high, the simulation is difficult to obtain accurate evolution. Generally, the model needs to be adjusted to adapt to these scenarios.

[0006] (2) Slow calculation speed, and real-time performance is difficult to guarantee

[0007] In order to obtain a more accurate evolutionary discharge model, a more complex physical model needs to be constructed. However, the calculation of these models is difficult to be fast based on the computing power of existing computers. The calculation time is hours or even days, and the real-time performance cannot be guaranteed, which seriously affects the real-time deployment.

[0008] (3) Failure to implement online detection

[0009] Tokamak magnetic probe fault diagnosis and detection based on existing technology cannot achieve real-time online detection, and is generally calibrated before the discharge experiment. When a fault occurs during the discharge process, the discharge will be interrupted and fail. After the discharge fails, researchers will evaluate and then calibrate to eliminate the fault, which seriously affects the progress of the experiment.

[0010] In summary, the present application proposes a real-time fault detection method for a tokamak magnetic probe. Summary of the invention

[0011] The purpose of the present invention is to propose a real-time fault detection method for a tokamak magnetic probe in view of the problems in the prior art that the fault diagnosis and detection of existing magnetic probes is difficult and slow.

[0012] The technical solution of the present invention is a real-time fault detection method for a tokamak magnetic probe, comprising the following steps:

[0013] Step 1: Collect neural network model training data to build a training database, and use the physical parameters related to the magnetic probe signal as model input;

[0014] Step 2: Using a neural network model with multiple encoding layers and decoding layers to capture the temporal evolution characteristics of the magnetic probe signal, and predicting the magnetic probe signal of the subsequent time period based on the input time period;

[0015] Step 3: transmitting the model input to the trained neural network model to predict the magnetic probe signal in real time;

[0016] Step 4: By comparing the actual measured magnetic probe signal with the predicted magnetic probe signal, it is determined whether there is a problem with the actual measured magnetic probe signal according to the set threshold value;

[0017] Step 5: Abnormal signal feedback processing: When a problem is detected in the actual measurement signal, the predicted signal will replace the problematic signal and output it to ensure the stability of the experiment.

[0018] Optionally, in step one, the physical parameters related to the magnetic probe signal include the voltage signals of the PF coil and IC coil and the total plasma current, and the state parameters of the current plasma, namely, the stored energy Wmhd, the ring voltage Vloop, the polar pressure ratio and the internal inductance Li, are used as inputs of the model.

[0019] Optionally, in the step of constructing a training database, the data source is the discharge data of the tokamak EAST device, including double-zero and single-zero discharge configurations, and the total current of the plasma in the top stage of the discharge covers the range of 200-500kA.

[0020] Optionally, the step 2 specifically includes: using the Transformer neural network model as the basic model to capture the temporal evolution characteristics, wherein the Transformer neural network model has 4 encoding layers and 6 decoding layers, the input is the 0-t time period, and the magnetic probe signal of the evolution time period from t to t+999 is predicted.

[0021] Optionally, the input of the Transformer neural network model includes plasma state parameters, which enter the Conv1D layer after Embedding and PE operations, and then undergo encoding and decoding processing through multiple layers of LN, Conv FFN, MHA and Masked MHA operations, and finally obtain a prediction signal through Projection and Linear operations.

[0022] Optionally, in step 4, the actual measurement signal is compared with the predicted signal. When the difference between the two is greater than a set threshold, it is considered that there is a problem with the actual measurement signal, otherwise it is normal;

[0023] Optionally, in step 4, the threshold in the step of comparing the actual measurement and predicted signals is set according to the actual operating environment and the performance requirements of the magnetic probe, and the threshold is set to 0.05, that is, 0.05.

[0024] Optionally, in the abnormal signal feedback processing, the output signal is marked after the problematic signal is replaced.

[0025] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0026] (1) An accurate plasma evolution model is constructed to predict the magnetic probe signal. The operating status of the magnetic probe is determined by comparing the actual measured value with the predicted value, thereby realizing magnetic probe fault detection.

[0027] (2) In terms of constructing the evolutionary model, the Transformer model is used to adapt the information containing time evolution and improve the accuracy of the model evolution.

[0028] (3) The GPU-based neural network evolution model overcomes the problem of long calculation time of traditional evolution models and realizes the rapid evolution of plasma fault diagnosis.

[0029] (4) The entire invention utilizes a data-driven artificial intelligence algorithm to replace traditional physical modeling, thereby improving the model prediction accuracy while also reducing the difficulty of model construction.

[0030] The present invention proposes to use artificial intelligence algorithms to realize fault detection of magnetic probe diagnosis in tokamak devices, which not only improves the calculation accuracy of the model, but also greatly optimizes the model calculation time, making real-time detection of magnetic probe faults feasible and having strong universality. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A principle block diagram of a real-time fault detection method for a tokamak magnetic probe is given;

[0032] Figure 2 This is a schematic diagram of the structure of the Transformer neural network model;

[0033] Figure 3 This is a graph showing the results of predicting 1000 steps of evolution based on the database-based training of the neural network model. DETAILED DESCRIPTION

[0034] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0035] Example

[0036] like Figure 1 As shown, the present invention proposes a real-time fault detection method for a Tokamak magnetic probe, and the specific steps are as follows:

[0037] (1) Collect neural network model training data and build a training database.

[0038] Based on the physical background, the contribution sources of the magnetic probe signal are fully considered, including the voltage signals of the PF coil and IC coil and the total plasma current, etc. Secondly, the state parameters of the current plasma, such as the stored energy Wmhd, the ring voltage Vloop, the poloidal pressure ratio and the internal inductance Li, are also considered as the input of the model.

[0039] (2) Build a neural network model and train the neural network model.

[0040] Since the present invention hopes to obtain a plasma evolution model, the model needs to consider the evolution over time, so the Transformer neural network model is used as the basic model to capture the evolution characteristics over time. The Transformer neural network model can effectively capture long-distance dependencies in the sequence, and can greatly shorten the training and reasoning cycle through parallel computing. Figure 2 As shown in Figure 1, there are 4 encoding layers and 6 decoding layers. The input of the model is the period from 0 to t ( Figure 2 The blue rectangular box in the middle) predicts the magnetic probe signal from evolution t to t+999 (the green rectangular box in the figure).

[0041] (3) Using the trained evolution model to predict magnetic probe signals in real time

[0042] The required inputs of the neural network model are transmitted to the neural network model, and the signal of the magnetic probe is predicted in real time.

[0043] (4) Comparison of actual measured and predicted signals

[0044] The actual measurement signal is compared with the predicted signal to determine the current operating status of the magnetic probe diagnosis. When the difference between the two is greater than the set threshold, it is considered that the actual measurement signal has a problem, otherwise it is normal.

[0045] (5) Abnormal signal feedback processing

[0046] When a problem is detected in the actual measurement signal, the predicted signal will replace the problematic signal and output it, thus ensuring the stability of the experiment.

[0047] Experimental verification

[0048] Based on the currently operating tokamak device, the discharge data from December 2022 to June 2024 were selected as the database. Among them, the discharge time exceeded 2 seconds, and about 15,000 shots (sampling rate 1kHz) of historical discharge data were screened out. These data include multiple discharge configurations such as double zero and single zero, and the total plasma current at the top of the discharge covers the range of 200 kA to 500 kA. Based on these databases, the neural network model was trained to predict the results of 1000 steps of evolution. Figure 3 As shown, the red one represents the actual measured magnetic probe signal, and the blue one is the predicted magnetic probe signal. It can be seen from the figure that the results of the evolution prediction are basically consistent with the actual measurement. The core of the present invention is the need for an accurate evolution model to achieve fault detection of the magnetic probe. This example can prove that the method based on the present invention can accurately realize evolution prediction, thereby realizing the monitoring of equipment failure. The model evolution prediction time in this example is about 1 ms, which verifies that the present invention has the ability of real-time monitoring.

[0049] The present invention proposes to use artificial intelligence algorithms to implement fault detection of magnetic probe diagnosis in tokamaks, which not only improves the calculation accuracy of the model, but also greatly optimizes the model calculation time, making real-time detection of magnetic probe faults feasible. In addition, the fault detection method proposed in the present invention (including establishing an evolutionary model based on data-driven) is not limited to magnetic probe diagnosis fault detection of tokamaks, but is also applicable to other diagnostic systems and even other fields that require real-time detection of the operating status of mechanical equipment, and has strong universality.

[0050] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A real-time fault detection method for a tokamak magnetic probe, characterized in that: The following steps are involved: Step 1: Collect neural network model training data to build a training database, and use the physical parameters related to the magnetic probe signal as model input. The physical parameters related to the magnetic probe signal include the voltage signal of the PF coil and the IC coil and the total plasma current. At the same time, the state parameters of the current plasma, such as the stored energy Wmhd, the ring voltage Vloop, the poloidal pressure ratio, and the internal inductance Li, are used as the input of the model. The data source is the discharge data of the Tokamak EAST device, including double zero and single zero discharge configurations, and the total current of the plasma in the top stage of the discharge covers the range of 200-500kA. Step 2: Use a neural network model with multiple encoding layers and decoding layers to capture the temporal evolution characteristics of the magnetic probe signal, and predict the magnetic probe signal of the subsequent time period based on the input time period, specifically including: using the Transformer neural network model as the basic model to capture the temporal evolution characteristics, wherein the Transformer neural network model has 4 encoding layers and 6 decoding layers, the input is the 0-t time period, and the magnetic probe signal of the evolution time period from t to t+999 is predicted. The input of the Transformer neural network model includes plasma state parameters, which enter the Conv1D layer after the Embedding and PE operations, and then undergo encoding and decoding processing through multiple layers of LN, Conv FFN, MHA and Masked MHA operations, and finally obtain the predicted signal through Projection and Linear operations; Step 3: transmitting the model input to the trained neural network model to predict the magnetic probe signal in real time; Step 4: By comparing the actual measured magnetic probe signal with the predicted magnetic probe signal, it is determined whether there is a problem with the actual measured magnetic probe signal according to the set threshold value; Step 5: Abnormal signal feedback processing: When it is detected that there is a problem with the actual measurement signal, the predicted signal will replace the problematic signal and output it.

2. A real-time fault detection method for a Tokamak magnetic probe according to claim 1, characterized in that: In the step 4, the actual measurement signal is compared with the predicted signal. When the difference between the two is greater than a set threshold, it is considered that there is a problem with the actual measurement signal, otherwise it is normal.

3. A real-time fault detection method for a Tokamak magnetic probe according to claim 2, characterized in that: In step 4, the threshold in the step of comparing the actual measurement with the predicted signal is set according to the actual operating environment and the performance requirements of the magnetic probe, and the threshold is set to 0.05, that is, 4. A real-time fault detection method for a Tokamak magnetic probe according to claim 1, characterized in that: In the abnormal signal feedback processing, after the problematic signal is replaced, the output signal is marked.

Images