Abstract
Measurement of locked mode (LM) is critical for research into Magnetohydrodynamic (MHD) instabilities and plasma disruption. To accurately calculate the amplitude and phase of the LM, the $n = 0$ pick-up must be extracted and subtracted from the detector signals. This paper presents a new method using Neural Networks (NNs) to predict the $n = 0$ pick-up ($B_{r,n=0}$) on LM detectors in the J-TEXT tokamak. A novel Power Multiple Time Scale (PMTS) approach was developed, demonstrating outstanding regression performance across multiple frequency ranges. The PMTS models fit the $B_{r,n=0}$ signal with minimal error in both time and frequency domains. This method enables the extraction of $n > 0$ pick-up generated by resonant magnetic perturbations (RMPs) using only a single LM detector, offering a path toward optimizing detector distribution in future fusion devices.
Introduction: The Challenge of Magnetic "Noise"
In tokamak physics, magnetic confinement is essential for maintaining high-temperature plasma. However, instabilities such as "tearing modes" can arise. When these modes stop rotating, they become Locked Modes (LM), which are primary precursors to major plasma disruptions. Detecting LMs is vital for device safety and pulse termination management.
Physicists use magnetic detectors like Mirnov probes and saddle loops to monitor these instabilities. However, these sensors are "polluted" by background magnetic fields—specifically the $n=0$ component—originating from the device's equilibrium fields and eddy currents in the vacuum vessel. Traditionally, this background is canceled by subtracting signals from pairs of detectors placed $180^\circ$ apart, requiring at least 4 to 8 sensors. In future reactors like ITER, where diagnostic space is at a premium, reducing this sensor count is a priority.
Methodology: Power Multiple Time Scale (PMTS) Neural Networks
The researchers at J-TEXT developed the Power Multiple Time Scale (PMTS) Neural Network to address the "memory effect" of background signals, particularly those from eddy currents. Unlike standard NNs that only process instantaneous data, PMTS incorporates historical data to capture long-term trends.
The PMTS architecture uses a power-law distribution for its "look-back" window (e.g., $t, t-\Delta t, t-\alpha \Delta t, t-\alpha^2 \Delta t, \dots$). This allows the model to capture low-frequency historical data without the computational burden of a massive input layer. The network consists of two hidden layers using Tan-Sigmoid activation functions. Input features include plasma current ($I_p$), toroidal field ($B_t$), loop voltage ($v_{loop}$), plasma displacement ($d_x, d_y$), and currents from equilibrium field coils (Vertical Field, Horizontal Field, and Ohmic Heating).
The models were trained on experimental data from the J-TEXT Tokamak, including field-only discharges (to isolate coil-eddy current relationships) and clean plasma discharges lacking MHD instabilities.
Key Results and Discussion
The PMTS approach demonstrated significant superiority over standard Multi-Time-Scale (MTS) models, particularly in the frequency domain where error was reduced by nearly four orders of magnitude. The models achieved a coefficient of determination ($R^2$) between 0.91 and 0.96.
In field-only tests, compensation error was less than 2%, while in complex plasma discharges, it remained below 8%. As shown in Figure 3, the predicted $B_{regression}$ (magenta) tracks the experimental $B_{LFS}$ (blue) with high fidelity, allowing for the precise extraction of the underlying plasma response.
One limitation is the initial data loss; the model requires a historical window to begin predictions, meaning the first few milliseconds of a shot cannot be processed. However, the ability to accurately predict the $n=0$ background at a specific location means researchers can now extract $n=1$ or $n=2$ mode signals using only one detector instead of four.
Conclusion
The PMTS neural network provides a robust, physics-informed method for extracting background magnetic signals in tokamaks. By leveraging non-linear time look-backs, the model effectively handles the complex decay of eddy currents. This advancement not only improves the precision of locked mode detection but also offers a significant opportunity to optimize and reduce the number of magnetic diagnostics required for future fusion reactors.
Summary updated