Deep learning based quench detection method for fusion device high temperature superconducting magnet

Abstract

The application discloses a fusion device high-temperature superconducting magnet quench detection method based on deep learning, which comprises the following steps: obtaining a magnet quench voltage noise signal, a magnet current signal, a plasma current signal and an eddy current signal from historical operation data of the fusion device, taking the signals as a data set to train a deep learning model; continuously and real-timely collecting the magnet current signal, the plasma current signal and the eddy current signal during the operation of the device, using the offline trained deep learning model to real-timely calculate the magnet quench voltage noise signal; deducting the real-timely calculated magnet quench voltage noise signal from the real-timely collected magnet quench voltage detection signal to obtain a real magnet quench voltage signal, and thus realizing accurate detection of the quench of the high-temperature superconducting magnet of the fusion device. With the strong learning ability of the deep learning, the method can timely and accurately detect the occurrence of the quench of the high-temperature superconducting magnet, and guarantee the safe and stable operation of the high-temperature superconducting magnet in the fusion device.

Description

Technical Field

[0001] The present invention belongs to the field of quench detection of fusion devices, and in particular relates to a deep learning-based quench detection method for high-temperature superconducting magnets in fusion devices. Background Art

[0002] The magnet system is a crucial component of a superconducting fusion device. It plays a crucial role in heating, confining, and controlling the plasma. A quench is a fault condition in a superconducting magnet, representing the transition from a superconducting state to a normal state. To prevent quenches in superconducting magnets, fusion devices maintain a certain margin during operation. However, quenches in superconducting magnets due to thermal, electromagnetic, and mechanical disturbances are unavoidable. If a quench in the magnet system is not detected in a timely manner, it can lead to serious accidents such as magnet system burnout. Therefore, quench detection in superconducting magnets is crucial for the safe operation of superconducting fusion devices.

[0003] For the rapidly alternating superconducting magnets used in fusion devices, compensating for the complex induced voltage noise in the quench voltage detection signal is a major challenge facing quench detection systems. Currently, the widely used method is to compensate for the secondary induced voltage using the same winding and a fixed coefficient. However, even after compensation, the induced voltage noise can still exceed the quench detection threshold, leading to false positives in the quench detection system.

[0004] As fusion device operating parameters increase, the operating environment of superconducting magnets becomes increasingly complex. Accurate quench detection of superconducting magnets is becoming increasingly important for future fusion devices with higher parameters and more complex electromagnetic noise. Especially for the use of high-temperature superconducting magnets in fusion devices, timely and accurate detection of quench occurrence is difficult.

[0005] Therefore, how to use the powerful modeling capabilities of deep learning models to model the linear and nonlinear relationships that affect the quench voltage detection signal from the complex electromagnetic environment of the fusion device to remove the influence of electromagnetic noise on quench discrimination is a technical problem that technicians in this field currently need to solve. Summary of the Invention

[0006] To solve the above technical problems, the present invention provides a deep learning-based quench detection method for a high-temperature superconducting magnet in a fusion device. The method can model the linear and nonlinear relationships that affect the quench voltage detection signal from the complex electromagnetic environment of the fusion device, thereby removing the influence of electromagnetic noise on the quench voltage detection signal and obtaining a quench voltage detection signal with a high signal-to-noise ratio.

[0007] To achieve the above objectives, the present invention provides a deep learning-based method for detecting quench in a high-temperature superconducting magnet of a fusion device, the method comprising the following steps:

[0008] Step 1: Obtain the magnet quench voltage noise signal, magnet current signal, plasma current signal, and eddy current signal from the historical operation data of the fusion device to form a training data set and train the deep learning model;

[0009] Step 2: By continuously sampling the signals of the same winding and the optical fiber current sensor, the real-time values ​​of the magnet quench voltage detection signal, the magnet current signal, the plasma current signal and the eddy current signal are obtained;

[0010] Step 3: Preprocessing the acquired real-time values ​​of the magnet quench voltage detection signal, magnet current signal, plasma current signal, and eddy current signal, including filtering and data normalization;

[0011] Step 4: Input the real-time values ​​of the pre-processed magnet current signal, plasma current signal, and eddy current signal into the trained deep learning model, and calculate and output the magnet quench voltage noise signal in real time;

[0012] Step 5: Deduct the output magnet quench voltage noise signal based on the preprocessed magnet quench voltage detection signal real-time value, and output the magnet quench voltage detection signal real value.

[0013] Furthermore, the step 1 includes:

[0014] Preprocess the historical operation data of the fusion device used for training, including filtering and data normalization;

[0015] Initialize the weights and biases of each layer in the convolutional neural network layer and the long short-term memory network layer to random numbers with mean 0 and variance 1;

[0016] The magnet current signal, plasma current signal, and eddy current signal in the preprocessed historical operation data of the fusion device pass through each layer of the deep learning model in sequence, and the output value of the deep learning model regarding the magnet quench voltage noise signal is output;

[0017] The error between the output value of the deep learning model regarding the magnet quench voltage noise signal and the magnet quench voltage noise signal in the preprocessed historical operation data of the fusion device is back-propagated to calculate the gradient, and the parameters of each layer in the deep learning model are updated to minimize the error, and then training for the next batch is continued.

[0018] Furthermore, the step 2 includes:

[0019] Sampling the stainless steel wire wound together with the magnet to obtain the magnet quench voltage detection signal;

[0020] The optical fiber current sensors installed on the magnet current bus, inside the vacuum chamber and outside the vacuum chamber are sampled to obtain magnet current signal, plasma current signal and eddy current signal respectively.

[0021] Furthermore, the step three includes:

[0022] A 20 Hz low-pass filter is used to filter the real-time values ​​of the acquired magnet quench voltage detection signal, magnet current signal, plasma current signal, and eddy current signal;

[0023] The real-time values ​​of the filtered magnet quench voltage detection signal, magnet current signal, plasma current signal, and eddy current signal are standardized, and the mean of the data is converted to 0 and the standard deviation is converted to 1. The following formula is used:

[0024] ,

[0025] in, represents the obtained standardized data, x represents the data to be standardized, μ is the mean, and σ is the standard deviation.

[0026] Furthermore, the step 4 includes:

[0027] The real-time values ​​of the preprocessed magnet current signal, plasma current signal, and eddy current signal are used as the input of the trained deep learning model. Each layer of the trained deep learning model is passed through in sequence to calculate the output value of the model, namely the magnet quench voltage noise signal.

[0028] Furthermore, the step five includes:

[0029] The real value of the magnet quench voltage detection signal after compensation is obtained by subtracting the output magnet quench voltage noise signal from the real value of the preprocessed magnet quench voltage detection signal for quench discrimination.

[0030] If the compensated quench voltage detection signal reaches the quench detection threshold, the current of the magnet is discharged and an alarm is issued;

[0031] If the compensated quench voltage detection signal does not reach the quench detection threshold, the deep learning model is continued to be used to compensate the quench voltage detection signal.

[0032] The beneficial effects of the present invention are:

[0033] The present invention adopts a self-designed deep learning model, takes magnet current, plasma current and eddy current signals as input, and uses them to perform secondary compensation on the quench voltage detection signal in the fusion device, so as to minimize the influence of electromagnetic induction noise and improve detection accuracy. The deep learning model can extract local features in the input sequence, and at the same time capture the long-term dependencies and temporal dynamic characteristics in the input sequence, thereby enhancing the ability to extract electromagnetic noise characteristic signals. The method can use the powerful learning ability of deep learning to take the quench voltage noise influencing factors as input, accurately calculate the quench voltage noise signal, and obtain a quench voltage detection signal with a high signal-to-noise ratio, thereby improving the signal-to-noise ratio and ensuring the safe and stable operation of the high-temperature superconducting magnet in the fusion device. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of the deep learning-based quench detection method for a high-temperature superconducting magnet in a fusion device of the present invention;

[0035] Figure 2 The circuit diagram for the same winding installation in the magnet;

[0036] Figure 3 This is a structural schematic diagram of the deep learning-based quench detection method for high-temperature superconducting magnets in a fusion device of the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0038] See also Figure 1 The deep learning-based quench detection method for a high-temperature superconducting magnet in a fusion device provided in an embodiment of the present invention may specifically include the following steps:

[0039] S10: Obtain the magnet quench voltage noise signal, magnet current signal, plasma current signal, and eddy current signal from the historical operation data of the fusion device to form a training data set, train the deep learning model, and achieve accurate calculation of the quench voltage noise signal;

[0040] The purpose of this example is to train a deep learning model. To adapt the model parameters to suppress voltage noise in quench voltage detection signals, the model is trained using historical data from the most recent 1,000 shots of the EAST fusion device. The parameters include the magnet quench voltage noise signal, magnet current signal, plasma current signal, and eddy current signal.

[0041] The model's goal is to calculate voltage noise from the input signal. Therefore, the magnet current, plasma current, and eddy current signals are used as inputs, and the magnet quench voltage noise signal is used as the output. By training on this thousand-shot data, the model learns the linear and nonlinear relationships between the input signal and the voltage noise.

[0042] The use of historical operation data of the fusion device to train the deep learning model includes:

[0043] The 1000-shot data used for training was filtered to remove high-frequency interference signals and then normalized to eliminate magnitude differences in the data.

[0044] To improve the convergence speed and performance of the model, the weights and biases of each layer of deep learning are initialized to 0 and random numbers with a variance of 1;

[0045] The input data passes through each layer of deep learning in turn to calculate the output value of the model, that is, the output value of the deep learning model regarding the magnet quench voltage noise signal;

[0046] The error between the output value calculated by the model and the actual quench voltage noise signal is back-propagated to calculate the gradient, and the parameters of each layer in the deep learning model are updated to minimize the error, and then training for the next batch is continued.

[0047] S20: acquiring real-time values ​​of a magnet quench voltage detection signal, a magnet current signal, a plasma current signal, and an eddy current signal by continuously sampling signals of the same winding and the optical fiber current sensor;

[0048] The purpose of this step is to obtain the signal required to compensate for the quench detection voltage in real time. The winding is made of a stainless steel wire embedded in glass fiber and installed at the four corners of the conductor. Figure 2 , Figure 2 This is a circuit diagram for a co-winding installation. One end of the co-winding is welded to the conductor at the end of the magnet, where the potential is V0. The other end is always wound parallel to the conductor. The arrow indicates the direction of current. Since the co-winding is always wound parallel to the conductor and in roughly the same position, the inductance of the magnet and the co-winding are roughly equal. Therefore, the potential difference between the co-winding and the magnet is V2-V1, that is, the co-winding signal is approximately equal to the quench voltage detection signal. .

[0049] However, due to the slight deviation between the position of the same winding and the center of the conductor it is wound with, the influence of various electromagnetic induction voltages cannot be completely eliminated, and signal compensation is required for the magnetic quench voltage detection signal obtained from the same winding.

[0050] The optical fiber current sensor is installed on the magnet current bus, inside the vacuum chamber and outside the vacuum chamber, and is used for real-time measurement of the magnet current signal, the plasma current signal and the eddy current signal, respectively.

[0051] S30: preprocessing the real-time values of the acquired magnet quench voltage detection signal, magnet current signal, plasma current signal and eddy current signal, including filtering and data standardization;

[0052] The purpose of this step is to preprocess the collected signals to improve signal quality. First, a 20Hz low-pass filter is used for filtering. The 20Hz low-pass filter is an electronic filter that allows low-frequency signals to pass through while blocking or attenuating signals above 20Hz frequency, which is used to remove high-frequency noise interference signals in the fusion device.

[0053] In order to eliminate the magnitude difference of the data and improve the training speed of the model, the collected data is standardized, the mean of the data is converted to 0, and the standard deviation is converted to 1, using the following formula:

[0054] ,

[0055] Where μ is the mean and σ is the standard deviation.

[0056] S40: inputting the real-time values of the magnet current signal, plasma current signal and eddy current signal after preprocessing into the trained deep learning model, and real-time calculating and outputting the magnet quench voltage noise signal;

[0057] The purpose of this step is to calculate the quench voltage noise signal. The real-time collected and preprocessed magnet current signal, plasma current signal and eddy current signal are used as the input of the deep learning model.

[0058] The input data passes through each layer of the deep learning in turn, and the output value of the model, i.e. the quench voltage noise signal, is calculated.

[0059] S50: based on the real-time value of the preprocessed magnet quench voltage detection signal, subtracting the output magnet quench voltage noise signal to output the real value of the magnet quench voltage detection signal.

[0060] In this step, the quench voltage noise signal calculated in the previous step is subtracted from the real-time collected magnet quench voltage detection signal to achieve the purpose of quench voltage detection signal compensation.

[0061] The compensated quench voltage detection signal is used for quench discrimination. If the quench voltage detection signal reaches the quench detection threshold, the current is rapidly discharged and an alarm is issued; if the quench voltage detection signal does not reach the quench detection threshold, the deep learning model is used to compensate the quench voltage detection signal.

[0062] Please refer to Figure 3 , Figure 3 The principle diagram of the fusion device high-temperature superconducting magnet quench detection method based on deep learning provided by the embodiment of the application. It contains an input layer, a data preprocessing layer, a convolutional neural network layer (convolutional layer and pooling layer), a long short-term memory network layer, and an output layer. The method takes the magnet current , the magnet current rate of change , the plasma current , the plasma circuit rate of change , the eddy current , the eddy current rate of change as the input of the input layer. After preprocessing by the data preprocessing layer, the features of the input data are extracted by the convolutional neural network layer (where the convolutional layer extracts local features, and the pooling layer prevents overfitting through dimension reduction). The feature sequence extracted by the convolutional neural network layer is processed by the long short-term memory network layer, thereby capturing the long-term dependence in the time series and enhancing the model's understanding ability of time series data, and calculating the quench voltage noise signal. Finally, the quench voltage detection signal is subtracted from the output of the long short-term memory network layer, i.e. the compensated quench voltage detection signal .

[0063] In summary, the application proposes a new type of fusion device high-temperature superconducting magnet quench detection method based on deep learning. The method uses a deep learning method combining convolutional deep learning and long short-term memory network. With the powerful modeling capability of deep learning, the input data features are extracted by convolutional deep learning, and the linear and nonlinear relationship between the input signal and the quench voltage noise signal is learned by the long short-term memory network layer. The induced voltage noise of the quench voltage detection signal is greatly reduced, the quench can be accurately detected, and false positives caused by peak noise can be effectively avoided.

[0064] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part. It should be noted that those skilled in the art can make some improvements and modifications to the application without departing from the principles of the application, and these improvements and modifications also fall within the protection scope of the claims of the application.

[0065] It should also be noted that, in this specification, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

Claims

1. A method for detecting quench in a high-temperature superconducting magnet of a fusion device based on deep learning, characterized in that: The steps include: Step 1: Obtain the magnet quench voltage noise signal, magnet current signal, plasma current signal, and eddy current signal from the historical operation data of the fusion device to form a training data set and train the deep learning model; Step 2: By continuously sampling the signals of the same winding and the optical fiber current sensor, the real-time values ​​of the magnet quench voltage detection signal, the magnet current signal, the plasma current signal and the eddy current signal are obtained; Step 3: Preprocessing the acquired real-time values ​​of the magnet quench voltage detection signal, magnet current signal, plasma current signal, and eddy current signal, including filtering and data normalization; Step 4: Input the real-time values ​​of the pre-processed magnet current signal, plasma current signal, and eddy current signal into the trained deep learning model, and calculate and output the magnet quench voltage noise signal in real time; Step 5: Deduct the output magnet quench voltage noise signal based on the preprocessed magnet quench voltage detection signal real-time value, and output the magnet quench voltage detection signal real value.

2. The method for detecting quench in a high-temperature superconducting magnet of a fusion device based on deep learning according to claim 1, characterized in that: The step one comprises: Preprocess the historical operation data of the fusion device used for training, including filtering and data normalization; Initialize the weights and biases of each layer in the convolutional neural network layer and the long short-term memory network layer to random numbers with mean 0 and variance 1; The magnet current signal, plasma current signal, and eddy current signal in the preprocessed historical operation data of the fusion device pass through each layer of the deep learning model in sequence, and the output value of the deep learning model regarding the magnet quench voltage noise signal is output; The error between the output value of the deep learning model regarding the magnet quench voltage noise signal and the magnet quench voltage noise signal in the preprocessed historical operation data of the fusion device is back-propagated to calculate the gradient, and the parameters of each layer in the deep learning model are updated to minimize the error, and then training for the next batch is continued.

3. The method for detecting quench in a high-temperature superconducting magnet of a fusion device based on deep learning according to claim 1, characterized in that: The second step includes: Sampling the stainless steel wire wound together with the magnet to obtain the magnet quench voltage detection signal; The optical fiber current sensors installed on the magnet current bus, inside the vacuum chamber and outside the vacuum chamber are sampled to obtain magnet current signal, plasma current signal and eddy current signal respectively.

4. The method for detecting quench in a high-temperature superconducting magnet of a fusion device based on deep learning according to claim 1, wherein: The step three includes: A 20 Hz low-pass filter is used to filter the real-time values ​​of the acquired magnet quench voltage detection signal, magnet current signal, plasma current signal, and eddy current signal; The real-time values ​​of the filtered magnet quench voltage detection signal, magnet current signal, plasma current signal, and eddy current signal are standardized, and the mean of the data is converted to 0 and the standard deviation is converted to 1. The following formula is used: ， in, represents the obtained standardized data, x represents the data to be standardized, μ is the mean, and σ is the standard deviation.

5. The method for detecting quench in a high-temperature superconducting magnet of a fusion device based on deep learning according to claim 1, wherein: The fourth step includes: The real-time values ​​of the preprocessed magnet current signal, plasma current signal, and eddy current signal are used as the input of the trained deep learning model. Each layer of the trained deep learning model is passed through in sequence to calculate the output value of the model, namely the magnet quench voltage noise signal.

6. The method for detecting quench in a high-temperature superconducting magnet of a fusion device based on deep learning according to claim 1, wherein: The step five includes: The real value of the magnet quench voltage detection signal after compensation is obtained by subtracting the output magnet quench voltage noise signal from the real value of the preprocessed magnet quench voltage detection signal for quench discrimination. If the compensated quench voltage detection signal reaches the quench detection threshold, the current of the magnet is discharged and an alarm is issued; If the compensated quench voltage detection signal does not reach the quench detection threshold, the deep learning model is continued to be used to compensate the quench voltage detection signal.

Images